Your SlideShare is downloading. ×
FINsights: Analytics in Collaboration with FICO
Upcoming SlideShare
Loading in...5

Thanks for flagging this SlideShare!

Oops! An error has occurred.

Saving this for later? Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime – even offline.
Text the download link to your phone
Standard text messaging rates apply

FINsights: Analytics in Collaboration with FICO


Published on

In the latest edition of FINsights, Infosys partners with FICO to walk you through the various issues and roadmaps of analytics with the goal of providing insights to help launch successful analytics …

In the latest edition of FINsights, Infosys partners with FICO to walk you through the various issues and roadmaps of analytics with the goal of providing insights to help launch successful analytics initiatives.

Published in: Economy & Finance, Business

1 Comment
No Downloads
Total Views
On Slideshare
From Embeds
Number of Embeds
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

No notes for slide


  • 1. Tap into the true valueof analyticsOrganize, analyze, and apply datato compete decisively
  • 2. PrefaceWelcome to the Analytics in Financial Services issue of FINsights!Analytics. As the Information Age advances, business and social discussions about analytics abound.Particularly among executives who are focused on profitable growth and risk management, theeffective use of analytics is increasingly viewed as critical to success.In response to growing demand from our clients for information and services related to analytics, wevededicated this issue of FINsights, Infosys thought leadership journal for the financial services industry, tothe topic of analytics. Indeed, this quarter we decided to do something special: weve merged the expertiseof Infosys and FICO to create a one-stop compendium of viewpoints, roadmaps and research piecesaddressing a topic which carries increasing importance in our data-drenched world.The last decade has been characterized by an explosion in the volume and complexity of information.Organizations have developed enormous data warehouses cataloguing everything from transactiondetails to online activity, to dates of birth. For financial institutions, this abundance of informationrepresents both a powerful opportunity and a daunting challenge. When effectively organized,analyzed, and acted upon, it can drive customer retention, reduce credit risk, and improve cross- andup-selling. However, arriving at the point of proper application requires significant knowledge of andinvestment in analytics—not only the science, but also the practice of applying it when facingtightened regulatory, fiscal and other real-world challenges.Though the potential of analytics is nearly unlimited, many organizations get caught up in the “whens”,“wheres”, “whats” and “hows”. This issue has been created as an enabler—a tool for our readers to tap intothe true value of analytics and fully realize the potential of the information available to them. In it, weaddress a variety of analytics trends and challenges which have emerged in recent years. From masteringthe intricacies of unstructured analytics, transaction analytics and adaptive analytics, to the governanceand management of information, this issue fuses the know-how of Infosys and FICO to help informyour analytics strategy.We hope you find the result of our partnership enlightening. Each article in this issue was forged froma unique combination of domain expertise, discussions with clients and analysts, exhaustive research,and inter-company reviews. Our contributors put a tremendous amount of time and thought intoeach article, with the goal of producing an issue of lasting value – one that you keep close at hand inthe months ahead. From ideation to publication, this issue has been exciting to develop, read, anddiscuss. We hope youll agree and look forward to hearing your thoughts, comments and suggestions.Ashok VemuriMember, Executive Council and Global Head, Banking & Capital Markets PracticeInfosys Technologies LimitedCharles illExecutive Vice President, Sales and MarketingFICO
  • 3. ContentPrefaceFrom the Editors’ DeskAnalytics for a New Decade01. Post-Crisis Analytics: Six Imperatives 0502. Structuring the Unstructured Data: The Convergence of 13 Structured and Unstructured AnalyticsRevitalize Risk Management03. Fusing Economic Forecasts with Credit Risk Analysis 2104. Unstructured Data Analytics for Enterprise Resilience 2905. Why Real-Time Risk Decisions Require Transaction Analytics 37Optimize to Drive Profits06. Ten Questions to Ask of Your Optimization Solution 4707. Practical Challenges of Portfolio Optimization 55Understand Your Customer08. Analytics in Cross Selling – A Retail Banking Perspective 6109. Analytics as a Solution for Attrition 6910. Customer Spend Analysis: Unlocking the True Value of a Transaction 77 011. A Dynamic 360 Dashboard: A Solution for Comprehensive 85 Customer UnderstandingFight Fraud More Effectively12. Developing a Smarter Solution for Card Fraud Protection 9313. Using Adaptive Analytics to Combat New Fraud Schemes 10314. To Fight Fraud, Connecting Decisions is a Must 109Improve Model Performance15. Productizing Analytic Innovation: The Quest for Quality, 117 Standardization and Technology GovernanceLeverage Analytics Across Lines of Business16. Analytics in Retail Banking: Why and How? 12517. Business Analytics in the Wealth Management Space 135
  • 4. From the Editor’s DeskIn 1992, Walmart launched the first terabyte database—the race for information was on. In the nearly20 years since, large financial institutions and retailers alike have piled up data at a seeminglyexponential rate. 100 terabyte databases are no longer the exception, they are the norm. Businessesacross industries have been remarkably effective at amassing information. Where they have struggled,however, is effectively and consistently utilizing this information to drive revenues and cut costs.To make this data actionable, having a strong analytics program is a competitive necessity. From fraudto risk to marketing to collections, financial institutions have been accelerating their use of analytics,moving from rearview analysis of historical data to forward-looking predictive analytics—models thatare predictive and drive optimal decisions. A recent Forrester survey found that 31% of IT decision-makers are implementing or planning to implement advanced analytics packages in the near future. Ifyour bank isnt investing, your competitors are.The information race of the last 20 years has paved the way for an analytics race over the next decade.Firms want to know with more certainty how their customers are going to behave—who is likely toattrite, where they will spend next, and who is most likely to default on a mortgage or credit card—andwhat actions will guide customer behavior in a mutually beneficial direction. How effectivelystrategies are developed and implemented to extract and leverage information will likely define thewinners and losers of the financial services industry in the 21st century. In this day and age,information is indeed power.This issue of FINsights has been created to help your organization tap into the true value of analytics.We start by looking forward, delving into the musts of a post-crisis business environment, thepotential of unstructured analytics, and the advances in credit risk analysis. From there, we provide aseries of roadmaps and recommendations for applying analytics to three of the major challengesbanks face today: fraud, improving understanding of the customer, and risk management. The issuewraps up with a series of articles addressing analytics optimization, model performance, and vertical-specific analytics. We hope this issue serves as not only an interesting read, but also as a tool inlaunching and transforming your analytics initiatives.Chisoo LyonsVice President,Analytic Science,FICOSrinivas PrabhalaDelivery Manager,Head BCM STAR Technology Group,Infosys Technologies Limited
  • 5. FINsights Editorial Board ASHOK VEMURI Member Executive Council, MOHIT JOSHI Senior Vice President and Global Head of Sales Head - Banking and Capital Markets Practice Banking and Capital Markets Practice Infosys Technologies Limited Infosys Technologies Limited RAJESH MENON LARS SKARI Partner Partner and Practice Manager Infosys Consulting Infosys Consulting BALAGOVIND KESAVAN CHISOO LYONS Head of Marketing Vice President Banking and Capital Markets Practice Analytic Science Infosys Technologies Limited FICO SRINIVAS PRABHALA Delivery Manager Head BCM STAR Technology Group Infosys Technologies Limited
  • 6. Analytics in Financial Services01Post-Crisis Analytics: Dr. Andrew Jennings Chief Research Officer and Head of FICO Labs,Six Imperatives FICOWhile todays business press is filled with the message that analytics can help companies“do business smarter”, the largest gains will come from using analytics smarter. FICO hasidentified six imperatives for analytics in an environment where the past may no longer bea good model for the future. At the core of these imperatives is the decision model—a workingmodel that explains the relationships between all the drivers of a decision and its results.The decision model can be used in business planning and is a critical element in strategyoptimization. Using such tools is essential to understanding the business situation andcreating lasting business advantage. More analytics does not mean more of the same, Introduction however. Lessons learned from the crisis must now shape the way organizations build andOne response to the severity of the recent use analytics. Grave dangers lie in the naïveeconomic crisis is an increasing demand application of pre-packaged analytic techniques.for analytics. Forrester Research predicts thatthe market for predictive analytics and data With that in mind, this article describes sixmining will grow at a rapid pace to US $1.8 ways analytics needs to change to drive businessbillion by 2014. performance to higher levels.
  • 7. buy a particular product, which related 1. Turn the “360-degree Customer view” Inside Out products are they most likely to buy within a specific range of time? Which channels does The new normal requires that companies will this customer use most, and does that pattern ultimately achieve integration and vary by season or by day of the week? coordination across multiple business lines, By using analytics to answer these kinds of products, channels and customer lifecycle questions, you can predict an individual management areas. At the same time, top customers sensitivity to the specific attributes performers are pushing beyond this effort to of an offer. You can also automatically generate bridge corporate views of the customer and population segments with similar sensitivity. focusing on developing a more holistic This is the key insight on which differentiation understanding of the customers view. In a can be built. sense, it is about “getting inside the customers head” and taking a 360-degree 2. Model the Decision to look outward. Optimize Performance To construct this inside-out customer view, While companies must continue to improve companies are bringing together data from scores and other predictive analytics, the a widening range of internal and external largest performance gains going forward will sources. The sense of urgency that came come from improving how analytics are used with the crisis is impelling them to tackle in decision strategies. It is, after all, better the thorny integration and organizational decisions that improve business performance, issues involved in sharing data across not better predictive models in isolation. product lines, channels and customer lifecycle decision areas. This means there is a need to model more than individual customers and their There is another aspect to this effort. No behavior. There is a need to model the one doubts that more data and more decision itself. Decision modeling is a relevant data lead to better models. fundamental technique for understanding Winners, however, wont just use that and improving decision strategies. data to build better models—they will use it A decision model, as shown in Figure 1, can to ask better questions. incorporate any number of predictive Take, for example, the way response analytics along with a multitude of other and propensity modeling has traditionally inputs. been done in marketing. Often what you All decision makers have a view of the way a are actually modeling is the offer you particular business situation operates just made—e.g., “how many customers in whether they recognize it formally or not. this new target population will respond to The great advantage of the decision model and accept this existing offer?” concept is that it makes that view explicit A smarter use of analytics is to model the and ties it to an objective, like maximizing offer you are about to make. That sales or profit. It also makes clear the decision means asking questions about which variables—like credit limit and interest rate, products individual customers are most or channel and price—and the constraints likely to buy next, and when they are most that need to be met, such as expected losses likely to make the purchase. If customers not exceeding some pre-determined value.6
  • 8. Simplified view of a decision model for a credit line increase Figure 1At the core of this model are predictions initial expectations, rather than simplyabout how customers will react to potential following a long list of metrics.actions. Generally, these are the models that n because there is an explicit Third,determine the effectiveness of the decisions objective, the decision model forms anthat get created. However, it is very common excellent basis for the comparison andfor decision makers and analysts to create simulation of one possible strategytheir decisions with a complete absence of against another.any formal understanding of these action-reaction relationships. Not surprisingly, this In addition to making the drivers of aleads to lost opportunity. decision clear, a decision model can be “solved” by mathematical optimization.The decision model concept creates the best Optimization pinpoints the single beststarting place for three crucial steps to strategy for maximizing or minimizing aimproving decisions: particular goal. The output of optimization isnadapting the parameters andFirst, the assignment, at the level of individualconstraints of the model for some new customers, of the best actions or treatmentscircumstance forms a sound way of for what youre trying to accomplish at thecreating new decision logic. You are portfolio or organizational level. The resultingaddressing the change at the structural solution can be expressed as a decision tree.level, not editing some derived construct. Using optimization and simulation with an because the decision modelSecond, well-developed decision model, you canmakes the underlying relationships explore how much impact a currentexplicit, it becomes much clearer what constraint is having on your projected resultsinformation should be tracked to and what would happen if you adjusted it. Forunderstand if a strategy is playing out as example, you could answer the question: “Ifexpected. You can track results against we allowed a slightly higher level of bad debt 7
  • 9. Optimizing a credit line management decision strategy Figure 2 losses, could we reduce attrition and improve only the economy, but changing customer profit?” As shown in Figure 2, this process behavior and new regulations. These identifies a spectrum of optimized developments are causing even more changes strategies—an “efficient frontier” of potential in consumer behavior. operating points. Exploring this frontier To cope with such upheavals, top performers helps you better understand key performance are continuing to analyze historical data, but dynamics and select the optimal operating also probing their markets by conducting point that is currently best for your business. rapid-cycle designed experiments. Theyre The drivers of an effective decision are there analyzing the results to learn as quickly as whether the decision maker recognizes them possible about changing market dynamics, explicitly or not. An analytically smart and to identify which offers, policies and business will realize this, and be well along the actions are working best now. They are path to improving business performance. following the proven principles of experimental design to learn more from fewer 3. Rethink What it Means to tests. This mature discipline aims at creating be Data-driven When the experiments whose results can be accurately Future is Not Like the Past analyzed and causes of variations understood. A well-designed series of To use analytics in smarter ways, companies experiments eliminates the need to test every can no longer rely solely on historical data. option—or even all the best options—because This can be seen clearly in the financial enough has been learned already to services industry, where consumers have extrapolate the outcome. changed their behavior in response to In a high-performing analytic organization, a economic stress. Creditors have changed certain proportion of new or “challenger” their offers and policies in response to not strategies will be purposefully designed to8
  • 10. produce “controlled variation.” If you test individual with a credit score of 620 is likelyonly challenger strategies that are close to to have a default rate much closer to thathow you currently do business, you will limit traditionally associated with a score of 614.what you can learn from your data. If, In a very severe recession, the default rate ishowever, you push the design of some likely to look a lot more like that of a 600.challengers outside of the bounds of business This does not mean that the score isas usual, you will introduce variation into “broken”—it may rank-order risk just asyour data and expand what you can learn strongly as ever, but the performance offrom it. customers at each score has degraded.In a practical situation, it is never as simple With this kind of economic impactas the challenger beats the champion. simulation, lenders can select the economicWell-designed strategies do not even set forecast they believe is probable, then use theout with this goal in mind. They set out to associated index metrics to adjust their scorepush the relationships that underlie the cutoffs for credit approval. In this way, it isdecision model so that learning increases possible to maintain a fairly consistentunderstanding. Think of them not as default rate across changing economicchallengers to replace a champion, but as conditions.learning strategies. This cycle, in turn, leadsto a new champion and another round of 5. Balance Automationtesting. By exploring a wider range of with Expertisepossibilities, you increase the chances ofdiscovering unique insights that might More analytics does not mean less need forlead to competitive differentiation. You are human expertise. Automated discovery ofalso less likely to be caught flat-footed when patterns and automation of modelingthe forces of change shake up the status quo. processes are important tools that can speed the process of improving decisions. The 4. Factor Macro-economic economic meltdown has shown, however, Forecasts into Analytics that letting the machines do all the thinking can lead to catastrophe.It is clear that abrupt economic changes, such Analytic expertise informed by deep domainas severe recessions, can cause actual customer knowledge is essential to building effectivebehavior to shift from what historical data predictive and decision models. Thissays it should be. In the recent downturn, expertise is indispensable for dealing withcredit default rates rose significantly above everything from data bias to inadequatethe rates historically associated with standard sample score ranges. Todays analytic announcements focus on theToday, there are new analytic methods for need to use analytics to crunch throughforecasting how such macro-economic petabytes of data. While this is true, it is onlyconditions are likely to impact customer part of the story. Given todays dynamicbehavior and change results. These analytics market conditions, businesses need analyticsare now being used to generate an index of to make decisions for which they do not havehow much default rates are likely to increase petabytes of data—or at least not petabytes ofor decrease under a range of economic relevant data. That is why you need to rely onforecasts. This index enables lenders to see, analytic experts to get the most predictivefor example, that in a moderate recession, an value from what you have to work with. 9
  • 11. Expertise makes the difference between analytic complexity must be justified. models that perform so-so and those that A senior manager should be able to get clear perform at a very high level. This answers to: “Why does this customer decision performance edge comes from the experts need to be so complicated? What value are ability to interpret nuances in the data we getting from it? What are we learning? in order to find the best predictive How is this going to make tomorrows characteristics for a desired performance decisions better than todays?” outcome. It comes from knowing how to Some complexity, of course, is good. It is validate models without “over-fitting” them how a company finds new niches of to the data they were developed from, and customers and tries to improve the decisions how to fine-tune models to a companys it makes about those customers. Much specific real-world business conditions. unnecessary complexity, however, can be Above all, the analysts understanding of the traced back to two root causes: 1) Trying to context in which a model will be used—and improve decisions by starting with the the decision it is supposed to improve—is decision tree rather than the decision model; absolutely critical. and 2) The inefficiencies of the decision tree as a means of representation. 6. Justify Complexity and Increase Transparency Editing a decision tree of any size is an inherently risky venture. This form of Businesses are operating in an increasingly representation is visually complicated, complex world, and it follows that analytics and therefore, working within it, one can must often be complex as well. Nevertheless, easily make mistakes. By comparison, the Comparison of two decision trees Figure 3 New strategy design tools simplify the visualization of complex strategies. This helps in the comparison of different strategies, such as in this example, which highlights the score cutoff change difference in two collection strategies.10
  • 12. underlying decision model can be much nodes, and it is not unusual to have treessimpler to understand and change. Edit the with many thousands of nodes, hence thedecision model, and you will have a better magnitude of the problem.grasp of the structural reasons for the changesyou are making, and thus can have Conclusionconfidence in the decision tree derived fromit, no matter how complex. In the post-crisis era, the companies thatWhen it is necessary to edit or “prune back” succeed with analytics will not be those thatthe resulting decision tree a bit, you need simply use more of them, they will be thosetools that help you navigate across it to focus that use them in smarter ways. Above all, toin on just the places requiring attention. You really succeed with analytics, you need toalso need tools that help you understand how understand the context of the data, themuch fidelity to the decision model you are operational context of the decision and thegiving up by making these edits. underlying business relationships that you are modeling. You just cannot do that byThe best tools also enable you to compare the going click-click-click.original decision strategy to the changedstrategy. Starting with the decision model,you understand the structural reasons for the Referenceschanges. Comparing the trees then shows youthe differences in the decision logic. This just 1. Market Overview: The Business Intelligenceis not possible when comparing “raw” trees Software Market, Forrester Research, Inc.with anything beyond a small number of 10/23/09, p7 11
  • 13. Analytics in Financial Services02Structuring the Unstructured Bala Venkatesh Group Project Kiran Kalmadi Senior Consultant, Shivani Aggarwal Consultant,Data: The Convergence of Manager, Infosys Technologies Infosys Technologies Limited Infosys Technologies LimitedStructured and Unstructured LimitedAnalyticsToday, 80% of business is carried out on unstructured data—documents, call center logs,blogs, wikis, tweets, and surveys. Neglecting to analyze such data leads to ignored risks,uninformed decisions, and missed opportunities. Financial services firms are increasinglyanalyzing unstructured data to understand customer needs, prevent frauds and expand thecustomer base. Analytics plays a key role in analyzing unstructured data and transforming itinto actionable intelligence. The rapid adoption of social media by the financial servicesindustry has resulted in an even higher percentage of unstructured data being generated.This has prompted firms to increasingly look at social analytics to derive structured insightsout of social media. As unstructured and structured data analytics are converging, financialinstitutions are looking for analytic vendors to come up with products that blendunstructured analytics (like social analytics) with structured analytics (risk analytics).This article analyzes unstructured data, the various analytics vendors in the space, andapplications in the financial services industry. have significant business value. However, due to Unstructured Data: What is it? insufficient search techniques and inadequate technologies, businesses are usually not ableUnstructured data refers to data that does to derive the right answers—leading tonot exist in a database. Unstructured data can inappropriate textual or non-textual and takes the formof text, audio and images (refer to Figure 1 for Unstructured analytics help businesses analyzemore on the different sources of unstructured unstructured data and transform it intodata). Unlike structured transaction data— actionable insights. These primarily consist ofwhich tells what customers did—unstructured text analytics, audio (or speech) analyticsdata provides insights into why they did it, and video (or image) analytics. Social mediawhat else they want to do, and what problems analytics are an important form of text analyticsthey may have. The answers to these questions making inroads.
  • 14. Types of data Figure 1 Structured Semi-Structured UnStructured ? & Legacy Relation XML ? Web ? Databases EDI Documents ? E-Mails ? Spreadsheets ? Wikipedia ? ? with Proper Flat Files Multimedia (Video, ? Record Formats Audio) RSS Feeds ? Within Corporate Messages ? Static Real Time Internal ? ? Center Customer Documents Call Logs Sales Report ? ? Center Customer Marketing ? Representative Notes Material ?Rooms Trading Analysts ? ? News Breaking Reports ?Pricing Market Internet Formal/Legal ? ? Events Weather Filings (SEC, ? Actions Corporate FDIC) Chat Rooms ? Journals ? Social Media ? Platforms Text Analytics Unstructured Data Analytics Text analytics enable businesses to derive Analysis of any form of unstructured value from large quantities of text. This text data that helps transform it into can be available either in existing repositories actionable intelligence is called or can be newly generated or acquired. This is “Unstructured Data Analytics”. Structured done by extracting and interpreting relevant data analytics uses business intelligence information to reveal patterns and relationships. Figure 2 elaborates the text tools for querying and reporting, analytics process in detail. whereas unstructured data analytics utilizes text processing and keyword Text analytics is gaining importance in all searches (to locate documents in servers). industry segments (specifically the financial Unstructured analytics has evolved over services industry), mainly because of the huge time, moving towards next generation chunks of textual data being generated techniques like video and audio analytics month after month by every organization (which are rarely used in the financial (both within and outside the organization). services industry) and text analytics With the advent of Web 2.0 & 3.0, there is a (also known as text mining). greater emphasis on information-sharing14
  • 15. Text analytics process Figure 2 Information Transforming Analytics Reporting Delivery Retrieval Text Collect and ? Content ? Selecting ? Different ? Taking steps to ? retrieve cleaning, attributes, mechanisms augment information removing discovering for notifying existing data & from both duplicates, patterns, results, like store enriched internal & language interpreting & dashboards, information external recognition, analyzing the alerts, etc. sources etc. resultsand user-collaboration using social In recent months, a host of social medianetworking sites. Interactions and posts on analytic products--from vendors such asthese sites play a huge role in shaping IBM, SAS, Scout Labs, and Radian6--haveconsumer sentiments about businesses, hit the markets. These products will helpservices, competition, and markets. Hence, banks and other financial services companiescorporations are investing in social media monitor and measure the performanceanalytics tools that use text analytics to of their social media campaigns—enablingunderstand customer sentiments, and banks to make informed decisions.address them proactively.Social Media Analytics Key Vendors in Unstructured Data AnalyticsSocial media analytics derive and measurekey results from social media. Social media The unstructured data analytics softwareanalytic tools use algorithms and approaches market is relatively new, and the vendors infor automated analysis of blogs, chats, emails this space are still emerging and shapingand other related social media. Key areas their lines-of-businesses. SPSS (acquired byaddressed by social media analytic tools are: IBM) and SAS are the major players. There are a number of other vendors whon the relevant blogs - IdentifyingFinding specialize in offering text analytic—tools andthe relevant blogs and forums for a tools for the monitoring and measurementbusiness of social media.nDetecting sentiment - Detecting the The Road Ahead for Vendorssentiments expressed about a company,product or new launch The future lies in bridging the gap between structured and unstructured data andnMeasuring the influence and authority creating products that blend unstructuredof key bloggers - Identifying the key analytics with structured analytics. Thebloggers and how they are influencing the challenge for vendors is creating structuredthought process of others on the web data from unstructured information. Then topics of interest masked inDetecting combination of various data sources andconversations on social network sites. different types of data to drive business is 15
  • 16. definitely worth exploring for vendors. and perspectives about a companys services The road ahead for vendors will be or products through social media. Most characterized by: financial services firms still rely only on structured data analytics for customer n Integration of social media analytics intelligence. Unless these structured data with text analytics- Vendors will be analytics are blended with social media looking to integrate social media analytics, it is very difficult to achieve analytics with text analytics. In recent actionable customer intelligence. times, social media monitoring vendors have started to merge with text mining Stock Market Prediction analytic vendors. Social media Predicting stock market movements is a monitoring as a stand-alone capability of challenge for investors due to lack of vendors will, thus, not stay for long. consistent prediction methods. However, n Text analytics will be made available research shows that there is a strong as a component of other applications- relationship between news stories and stock Text analytics applications will move price movements. Predicting the stock price away from being siloed applications movements based on news items is gaining towards being a part or a component of increased importance in the text mining other applications in the business. In community. such a set-up, insights derived from Fraud Detection analysis of unstructured textual content will automatically flow on a real-time Financial institutions lose millions of dollars basis into the business for key decision- to fraud every year. In banking, fraud arises as making. a result of stolen credit cards, forged checks, misleading accounting practices, etc. n Convergence of various types of analytics- There will be a rapid growth of Financial services firms need to have combinations of different types of improved analytical capabilities to reduce analytics. For instance, text analytics fraud levels and the associated costs. with predictive analytics is expected to A common thread in the above three make rapid headway. Financial services application scenarios is a proper blend of firms will begin using a combination of structured analytics and various forms of text analytics and predictive analytics for unstructured analytics. A well-blended risk management and fraud management solution is much better than traditional on a large scale. analytics solutions. A key challenge in using unstructured Applications of Unstructured Analytics in Banking, Financial data analytics is that the unstructured Services and Insurance (BFSI) data rarely has a consistent internal infrastructure, or metadata (unlike the Figure 3 highlights a few of the specific structured data), and hence it is far more application areas within the BFSI segment complex to analyze and model. Despite these where unstructured analytics can be readily difficulties, businesses are incorporating, deployed. and should continue to incorporate, unstructured analytics efficiently into their Customer Relationship Management processes, mainly because of the extensive Customers share ideas, insights, experiences business value derived from such analysis.16
  • 17. Application areas of unstructured analytics Figure 3 Customer Relationship Management Fraud Detection Stock Market PredictionImproving customer experience in a bank using unstructured analyticsScenario:noffered several services to customers through its online banking channel. EachA bank typically involved multiple complex user interactions.n using these services felt that the website was not user friendly and they startedCustomers moving to other banks based on that sentiment. They also started sharing negative sentiments about the bank in online forums.nMany customers did not even have the patience to respond to the online customer survey feedback form. Challenges Faced by the Bank Solution Recommendation using Unstructured Analytics ¡ Lack of tools for achieving a 360 ? Optimum Solution: A blend of view of customers, resulting in poor structured analytics using business customer retention rate. intelligence tools and social media analytics. Public relations disaster due to the ? ? utilizes social media analytic The bank negative sentiments expressed in tools to analyze online consumer online forums. forums and blogs. Along with this, the bank analyzes the Voice of Customer results collected from its own Business Intelligence (BI) tool. This blended solution helps to ? accurately predict the customer sentiment and invest intelligently on customer experience management. This a shift from the traditional approach of only analyzing the Voice of Customer survey and Net Promoter Score, which would not have yielded the desired result. 17
  • 18. Stock market prediction using unstructured analytics Scenario: Stock market research is primarily based on two trading philosophies, namely, a) Fundamental ? - in which the prediction is based on the securitys data—price to earnings ratios, return on equities, etc., and b) Technical - which uses charts and modeling techniques for prediction. ? fundamental and technical analyses, information from quarterly reports and Apart from breaking news also plays a major role in the movement of share price. Prediction has always been a challenge, since there has not been much success in analyzing this textual data. Challenges faced Solution Recommendation using Unstructured Analytics ? stock market prediction Inaccurate Optimum Solution: A combination of when a breaking news story or Structured analytics (using technical quarterly results are declared, leading approach) and Unstructured analytics to huge losses for the brokers and (using fundamental approach). investors. In a technical approach, the historical ? data of a stock is analyzed and a linear regression is run to determine the price trend. In a fundamental approach, article ? terms in financial news, shareholders reports etc. are assigned a weight. Using the Bag of Words, Noun Phrasing & Named Entities techniques*, the textual key words that relate to “earnings” or “loss” are identified. ? of both the approaches are The results combined to arrive at a predictable outcome (up, down, or unchanged movement of the stock price). ? 4 for understanding the Refer Figure solution steps in a typical financial news analytical system. Financial news analytical system Figure 4 Fundamental Approach Technical Approach (Unstructured Analytics) (Structured Analytics) Textual Analytics New Regression Stock Techniques like DB Article Analysis Quotes Bags of Words Stock Error Analysis Stock Market Market Prediction Model Prediction * These are specific approaches that use linguistic textual representations.18
  • 19. Detection of fraud in insurance companies using unstructured analytics Scenario: ? the insurance industry can occur in any stage of the transaction and can be Frauds in committed by any party (i.e. new customers, policy holders, third party claimants or any other party involved in the transaction). Typical frauds include inflating actual claims, misrepresenting facts, and submitting claims ? for damages that never occurred. Challenges Faced by Insurance Solution Recommendation using Companies Unstructured Analytics ?approach for fraud detection ·Reactive Optimum Solution: A combination of whereby the insurance companies structured predictive analytics, speech investigate only after a fraudulent analytics and social media analytics. claim is made. ? modeling combined with text Predictive ·Fraud investigation process is lengthy ? and social media analytics can be used and expensive, as an investigative to detect and prevent fraud. officer has to investigate personally to detect any suspicious activity in ? the predictive model are from Inputs to claims. internal & external watch lists of criminals who previously engaged in Limited fraud prevention mechanisms. ? fraud, diagnostic fraud indicators based on surveys taken by the claim handlers, anomaly patterns, profile details of individuals, etc. ? of the predictive model The results combined with the result of claimants speech analytics (to detect whether the claimant is lying or not) along with the results of social media analytics (to detect the relationship among the policy holders) can be used to generate a risk score of the claimant. ?score can then be used to This risk detect fraud customers even before issuing the policy. coming up with solutions to integrate Conclusion structured analytics with unstructured analytics. By converging unstructured analyticsThere is an increasing requirement into the structured analytics mix, businesseswithin organizations to inquire and are seeing substantial improvements in the accuracy and relevance of their analyticanalyze across structured and unstructured initiatives. Incorporating analytics blendsdata. Integrating unstructured data with into business processes is a growing trend;structured data is quite a challenge. Despite however, they must be correctly applied to athe challenges, many analytics vendors are specific business scenario, and companies 19
  • 20. must act on the results appropriately. The convergence of unstructured analytics with structured analytics is no longer an “if”, but rather a “when”.20
  • 21. Analytics in Financial Services03Fusing Economic Forecasts Dr. Andrew Jennings Chief Research Officer Carolyn Wang Senior Manager, and Head of FICO Labs, Analytics,with Credit Risk Analysis FICO FICOAs many in the financial services industry prepare for measured economic recovery, itscritical to take stock of a key lesson of the financial crisis: that risk, by its nature, is dynamic.Todays economic realities call for a paradigm shift in risk management—one that includesnew analytics that go beyond the traditional assumption that past risk levels are indicative offuture risk. This article discusses new methods for systematically incorporating economicdata into scoring systems, allowing lenders to balance consumer level information withchanging economic trends. It also shares results from lender applications of this methodologyto “get ahead of the curve” by more closely aligning risk strategies with future performance. In 2005 and 2006, for instance, a default rate of Why a New Approach is Needed 2% was associated with a score of approximately 650–660. By 2007, this default rate was associatedTraditional assumptions that recent default rates with a score of about 710—a 50 to 60 point shift.will be representative of future defaults works This risk shift was most likely due to economicreasonably well if the lending environment decline after a period of more lenient lendingremains relatively stable. But in situations where practices, such as adjustable rate mortgagesexternal factors are changing rapidly, it can be and no-documentation (“no-doc”) loans.dangerous to assume the risk levels associated Lenders in 2007 who made decisions assumingwith scores will remain stable over time. that 670 still represented 2% default saw increasesTake the recent credit crisis as a case in point. in their default levels, when the actual defaultFigure 1 (next page) shows the observed default was about 4.5%. ®rate at each FICO Score range for real estate loans. Many lenders do attempt to anticipate changesEach line represents a large random sample of in economic conditions and adjust strategiesexisting accounts evaluated over different time accordingly—for example, tightening originationperiods and the default rates one year later. policies and reducing credit lines in a recession,As the graph shows, the risk levels associated or loosening policies in an upturn. But lenderswith accounts in 2005 (blue curve) versus 2006 often make these changes judgmentally, and as a(red curve) were already diverging. One year later, result, there is a tendency to over-correct and missthere were much greater risk levels associated key revenue opportunities, or under-correct andwith all but the very highest scores (green curve retain more portfolio risk than desired.shifts more dramatically upward).
  • 22. Risk levels can shift dramatically over time Figure 1 Next Evolution of influence the risk profile of newly booked Predictive Analytics and existing accounts. n Competition. As consumers are To restore profitability in a post-crisis presented with more (or less) attractive economy, lenders need an objective and offers by the competition, attrition will empirical approach to measure changing change the population. conditions and translate these into more effective strategies. The next generation More specifically, lenders need analytics of predictive analytics must go beyond that predict the impact of the above factors the assumption that only past risk levels on future risk levels for each account. In other are representative of future risk. Instead, words, the analytics must consider the risk estimates should also account for the macro-economic view of market conditions impact of: within the micro-analysis of individual consumer risk. The analytics must be n The economy. Consumers ability to flexible enough to take into account what repay changes as the economy shifts. is known—the historical data available Some lower-risk consumers may to derive past patterns and current economic refinance in downturns, leaving behind conditions—as well as what is expected— a portfolio of riskier consumers. forecasted views of the future. Others may reach their breaking points through job loss or increased payment Given the ready availability of economic data, lenders can first take into account the impact requirements. Higher-risk consumers of economic factors on future risk estimates. get stretched further, resulting in more In addition, lenders should begin to track frequent and severe delinquencies and changes in their strategies and competitive defaults. factors, which can be incorporated into n strategies. Changes in lender Lender future risk estimates as the data becomes policies across the account lifecycle can robust enough.22
  • 23. Analytics Tuned to Future are expected to behave differently under Performance varying economic conditions. This methodology builds upon existingNext-generation analytics can provide risk tools used by lenders, enabling quicklenders with an understanding of how and seamless implementation. It can bethe future risk level associated with scores applied to a variety of scores, such aswill change, based on current and projected origination scores, behavior scores, broad-economic conditions. Based on past based bureau scores like the FICO® Score,dynamics, the analytics derive the and Basel II risk metrics.empirical relationship between the defaultrates observed at different score ranges When applied to these scores, lenders gain(e.g., the risk scores odds-to-score an additional dimension to their riskrelationship) as seen on the lenders predictions so they can better:portfolio, and historical changes in n Limit losses. Lenders would have greatereconomic conditions. Using this derived insight on how to tighten up creditrelationship, lenders can then input current policies sooner and for the rightand anticipated economic conditions to populations during a downturn.project the expected odds-to-score outcomeunder those conditions. They can examine n Grow portfolios competitively. Lenderseconomic indicators such as unemployment could more quickly determine when andrate, interest rate and Gross Domestic how to proactively loosen up creditProduct (GDP). policies as markets recover.With such a relationship, it is possible to n for the future. Lenders could Preparerelate the impact of economic factors on simulate the impact of future macro-odds, default rates and scores. This relationship economic conditions on scores,can be derived at an overall portfolio level or to better adjust longer-term strategiesmore finely for key customer segments that and stress-test portfolios. More accurately predict default rates with analytics Figure 2 tuned to economic impact 23
  • 24. n Meet regulatory compliance. Lenders Over a three-year time span, predictions ® could better set capital reserves by from the FICO Economic Impact Service 2 creating more accurate, forward-looking, reduced the error rate by 73%, compared to long-run/downturn estimates required the traditional prediction. by Basel II. In practice, economically adjusted Grow Portfolios Responsibly analytics generate a more accurate risk prediction thats better tuned to Within originations, the lender primarily economic conditions. For instance, FICO determines whether or not to accept the loan recently applied this methodology—called as well as the initial loan price/amount. Using ® the FICO Economic Impact Service—to economically adjusted analytics, a lender can a behavior score used by a leading US set a cutoff score based on the anticipated credit card issuer. Figure 2 (previous page) future default rate, as opposed to the historical compares three metrics: default rate. The lender can maintain the desired portfolio risk levels by adjusting cutoff n The actual bad rate observed on the scores as the economy changes. portfolio over time (blue line). Figure 3 shows how a lender can view the n rates predicted by the behavior The bad current default rate by score range (orange score aligned to historical odds line) and predict how the default curve would performance (orange line)—traditional 1 shift under different economic conditions. approach . During a recession, the curve may shift to the n rates predicted by the behavior The bad dark blue line. The lender can update its score aligned to anticipated odds cutoff score from the “Current Cutoff” to the performance (green line)—FICO ® “Cutoff–Recession” based on empirical Economic Impact Service approach. The guidance. This allows the lender to limit economic conditions used were limited losses by proactively tightening credit in to what was known at the time of scoring. anticipation of the downturn. Flexibly adjust score cutoffs under different economic conditions Figure 324
  • 25. Conversely, during a time of economic traditional behavior score across the fullgrowth, the curve may shift to the light range of account management line. The lender can update its cutoff As an example, for a US credit card issuer,score to the “Cutoff–Growth.” This allows FICO retroactively used a FICO® Economicthe lender to loosen credit policies in Impact behavior score in place of theanticipation of an economic upturn, and traditional FICO ® TRIAD ® Customerbring in more profitable customers ahead Manager pooled behavior score for a credit lineof competitors. decrease strategy and for collection actions. TheFICO recently partnered with a European key question was: in April 2008 (a period of ® ®lender to apply the FICO Economic relative economic calm), could FICOImpact Service to its personal loan portfolio. Economic Impact Service help the lenderFaced with sky-rocketing delinquencies anticipate the economic turmoil six monthsthat were 2-3 times historical levels, the later and minimize its financial impact?lender sought to beat the market by FICO analyzed performance in October 2008proactively adjusting origination strategies and compared the different decisions madeahead of its competition. Using the by the two scores. Figure 4 shows the lineresults from the economically adjusted decrease, the lender saw opportunity toimprove its profit per applicant by US $11.50. The columns of interest are the “Swap In” and “Swap Out”, since they illustrate whereA similar approach can be taken for initial different decisions would be made. Theloan amount and pricing strategies. second column identifies accounts that would have received decreases using the Improve Account Management Decisions Economic Impact score, but did not receive decreases by the traditional behaviorLenders use behavior scores to help score (the accounts would be “swapped in”manage accounts already on their books if the lender had used the Economicfor credit line management, authorizations, Impact score). The third column identifiesloan re-pricing and cross-sell decisions. accounts that the Economic Impact scoreAn “economically impacted” behavior score would not have decreased and the traditionalcould be used in place of or along with the behavior score did decrease. Economic Impact score better identifies higher-risk Figure 4 accounts for line decreases 25
  • 26. The highlighted cells show that the behavior also would have not decreased accounts less score for these two populations are almost the sensitive to the downturn, reflected by same (swap-in: 643 vs. swap-out: 646). In other slightly higher scores. words, the behavior score identified both The actual bad rates seen six months later populations as at relatively the same risk level. reinforces that the Economic Impact score ® However, the FICO Economic Impact score identified riskier accounts (swap-in: 10.5% vs. was better able to distinguish risk among swap-out: 7.9%). If the lender had decreased these populations. Using this score, the credit lines on the appropriate accounts, it lender would have decreased more accounts could have realized a yearly loss savings of that would be negatively affected by the roughly US $2.4 million and a net savings of downturn (average score of 625). The lender US $1.7 million, shown in Figures 5–6. ® Yearly loss savings using FICO Economic Impact Service Figure 5 ® Yearly net savings using FICO Economic Impact Service Figure 626
  • 27. Figure 7 illustrates how the lender could Limit Collection Losses have saved close to US $4 million by taking aggressive action earlier. FICOIn times of economic turmoil, its even more calculated this using the number of actualcritical for lenders to proactively manage bad accounts that would have receivedcollection efforts. Economically adjusted accelerated treatment, average accountscores can be applied on existing behavior or balance and industry roll rates.collection scores to help lenders identifywhich accounts will become riskier and Combining this roughly US $4 million inshould receive increased collection priority. collection savings with the US $1.7 million savings from the credit line decrease strategy,For the same US card issuer, FICO the lender would have saved US $5.6 million.retroactively used an economically impacted This illustrates the aggregate benefits of thebehavior score in place of the traditional service when used across two areas in abehavior score to treat early-stage (cycle 1) customer lifecycle. Clearly, the benefitsdelinquent accounts. Prioritizing accounts by would be scalable for larger portfolios. ®risk, the strategy using the FICO EconomicImpact behavior score would have targeted Set more Accurate Provisions41% of the population for more aggressive and Capital Reservestreatment in April 2008. FICO thenexamined the resulting bad rates six months When setting provisions and capital reserves,later (October 2008), and saw that these it is important to understand the risk in theaccounts resulted in higher default rates. portfolio under stressed economic conditions. Having forward-looking riskIn other words, the Economic Impact score predictions is explicitly mandated by Basel IIbetter identified accounts that should receive regulations, and should be part of anymore aggressive treatment in anticipation of lenders best practice risk management.the downturn six months later. Using thisstrategy, the lender would have been ahead of FICO worked with an Eastern Europeanits competition in collecting on the same lender to apply FICO® Economic Impactlimited dollars. Service to its Basel II Probability of Default ® Yearly loss reduction using FICO Economic Impact Service Figure 7 27
  • 28. (PD) models. Using the derived odds-to-score aligned to current and future expected relationship between its PD score and economic conditions, lenders can more economic conditions, the lender can quickly adjust to a dynamic market and steer simulate the expected PD at a risk grade their portfolios for the uncertainties ahead. level under various economic scenarios. Thus, the lender can more accurately References calculate forward-looking, long-run PD estimates to better meet regulatory 1. Some scores are periodically “aligned” to requirements and calculate capital reserves. maintain a consistent odds-to-score relationship over time—for example, to Redefine Risk Management ensure a behavior score of 675 equals a Best Practices target odds of 30 to 1. Traditionally, behavior scores have been aligned to the Theres no better time for lenders to re- odds observed in the last six months. evaluate risk management practices in order to better prepare for measured growth or 2. Error rate is defined as the absolute buffer against a lingering recession. Forward- difference between the actual bad rate and looking analytic tools will become the risk predicted bad rate as a percentage of management best practices of tomorrow. actual bad rate. With improved risk predictions better28
  • 29. Analytics in Financial Services04Unstructured Data Dilip Nair Project Manager, Srinivasan V Ramanujam Engagement Manager, Allen Selvaraj Senior Technical Architect, Banking and Capital Banking and Capital Systems IntegrationAnalytics for Enterprise Markets Practice, Markets Practice, Practice, Infosys Technologies Infosys Technologies Infosys TechnologiesResilience Limited Limited LimitedThe recent financial crisis has shown us how various industries and entities are linked to eachother through a complex web of relationships. This crisis also exposed the ability (or lackthereof) of companies, industries and even countries to plan and respond to the changesaround them. This article talks about how an analytics platform can help organizations tomonitor and manage changes proactively—thereby infusing a dose of resilience against rapid,unexpected changes. Two large financial organizations, Bank-A and its Introduction primary competitor, Bank-B, relied on the same offshore Vendor-X. When Vendor-X had financialEnterprise Resilience is an effort across the troubles, resulting in huge attrition and filing oforganization to anticipate and successfully bankruptcy, these two banks were at risk ofnavigate adversity. Literally, resilience refers to experiencing potential disruption of ongoingthe power or ability to regain the original shape, projects.form or position after being bent, compressed orstretched (subjected to adverse conditions). The Bank-A identified and responded to this incidentterm Enterprise Resilience has been subject to in real-time; they worked with other vendors tovarious interpretations in the industry as it varies ensure that their projects were de-risked beforefrom government agencies, to non-government the actual bankruptcy of Vendor-X occurred.organizations, to financial organizations, to IT Bank-B, on the other hand, was slow to identifyvendors. Enterprise Resilience is not disaster and respond to the incident. Their projectsrecovery or business continuity, but is the new continued to depend on Vendor-X, withDNA for risk management that alters the insufficient back-up, because Bank-A had theorganizations approach from proactive to first movers advantage to scoop up availableadaptive, assuming that the bridge from reactive resources. What followed is predictable. Bank-Asto proactive has been crossed. projects were delivered on time and under budget, and they were able to gain market share,Consider the following hypothetical example. It whereas Bank-B lost market based on a real life scenario involving twoorganizations encountering the same incident, While this is a hypothetical scenario, the realitybut responding very differently. of market share gained and lost defines the new
  • 30. world of business. In this flat world, the reported their 20 t h billion tweet). definition of risk, dependencies, and Another study puts the value of social risk management have all changed. networking sites beyond the Gross Domestic Traditionally, risk management delivered Product (GDP) of some countries (Facebook one-dimensional solutions—focused on – 11.5 billion, Twitter – 1.4 billion and mitigating risks by addressing vulnerable growing). This valuation is undoubtedly tied areas. This approach has failed to include to the enormous amounts of unstructured and address the various interdependencies data that users create on their profile pages, that characterize the flat world we operate walls or tweets. in today. By understanding the broader The best way to be prepared for any threat risk, and managing risk across the is for an organizations operation to mine extended enterprise, Bank-A demonstrated this huge amount of data and extract greater Enterprise Resilience. meaningful scenarios. In this article, the With globalization, enhanced regulatory focus is on how to mine unstructured scrutiny, a competitive marketplace, data channels with a specific objective in mind—build better resilience for an organizational interdependencies (at a organization. For simplicity, two data level never seen before), and evolving channels will be featured in this article technological challenges emerging, the types (highlighted section in Figure 1). The of risks being faced by the organization concept can be replicated for most other are new, unique and challenging. Being information sources: proactive and adaptive requires that organizations be aware of the ever-changing n (non-formal): With a strong user- · Twitter business and geo-political environment. base, any breaking news is immediately Doing so requires that organizations harness propagated through tweets. (Tweets are every bit of data that is available—federal text-based posts of up to 140 characters and regulatory updates, geo-political news, displayed on the authors profile that are weather updates, emergency service updates, visible either publicly by default or by business and financial updates, and social restricted followers.) networking sites. There are several services nnews websites (formal): Websites— · Media that offer the above in a packaged format. such as NewzCrawler, FeedDemon, However, one of the most undertutilized Google Reader, Omea Reader, Bloglines, data sources is the unstructured data NewsGator and scoopitonline—act as emerging from social networking sites. collectors of breaking news from various According to data scientists, nearly formal news channels and RSS feeds. 80% of todays data is unstructured. New information channels—like the Mining/ Modeling – Identification and internet (Facebook, Twitter), email, instant Classification messaging (chat), text messaging (SMS), and voice-over-IP (phone calls)—are The data channels described in Figure 1 are generating enormous stores of non- unstructured. The next step in the process traditional data at a mind-boggling is to mine/ model this data to derive potential pace. According to recent estimates, Twitter scenarios (highlighted section in Figure 2). users create approximately 7 Terabytes of There are many unstructured data analytics data in a day (on 30th July 2010, Twitter packages available in the market (for example,30
  • 31. Step 1: Twitter and media websites as data channels Figure 1 Rich Profile Data (Location Attributes and Past Events) Twitter RS S Fe ed Improved s Data Google Mash-ups Analytics Response Landing Text Store for Geo-locational Reports & Zone Analytics Engine Times & Cost (ODS) Tagging Dashboards Savings Identification s eed & Classification SF Media RS WebsitesHIMI from Infosys, AeroText from Rock perform sentiment analyses to enablesoftware, Attensity360 from Attensity, behavioral and predictive capabilities.Lexalytics, Enterprise Miner data-mining n · Categorization of data and classifyingworkbench from SAS). These packages it into appropriate buckets, likeprovide a very good GUI-based approach environmental disasters, man-madefor viewing data to build, test and publish incidents/ threats, stock market changes,models using: geo-political activities like administrationn for words, word patterns or· Searches changes, policy changes, civil unrest, strikes strings to identify emerging events and to and boycotts, and other forms of gain insights into potential threats. The disruption to business. input to this would be a library of words, n fine-tuned by client, industry · Patterns word patterns or strings that are pre- segment and geography. The objective is selected and fed to the tool. to vary these patterns so they can capturen· Distinguish message patterns (tweets) for small, medium or large incidents while emotions and stress. This can be used to avoiding false positives. Step 2: Identification and classification of incoming data Figure 2 Rich Profile Data (Location Attributes and Past Events) Twitter RS S Fe ed Improved s Data Google Mash-ups Response Landing Text Store for Geo-locational Analytics Reports & Zone Analytics Engine Times & Cost (ODS) Tagging Dashboards Savings Identification s d Fee Media & Classification S RS Websites 31
  • 32. During this process, one must pay close The first step towards building a robust attention to data completeness, data quality analytics engine is to source and seed and data standardization. Missing data ancillary data to enhance and enrich data points should also be addressed—enabling profiles. This can be done by attaching data filling for good hypothesis creation. attributes to the organizations location parameters, such as: Geographical Tagging n functions of an organization Business and Mapping and their geographical parameters Next in the process, is the geographical n indicators Key risk mapping of the identified messages n of employees working at various Number (highlighted section in Figure 3). Thanks locations to tweet stream RSS feeds that include geo-location data the tweets can easily be n and value of all assets and Number tagged to specific geographic coordinates via resources deployed at the location the Google Maps APIs. Similar tagging n Competitor information can also be done for media website RSS feeds that, by default, carry the location The data is then analyzed for: information. n creation Scenario Once the messages are tagged to n of scenarios Persistence specific geographic coordinates, they will then be mapped to the coordinates of n Impact Analyses Cause and an organization and its competitors Time dimension and related information locations. can be brought in for trending reports and historical data analyses. Historical Analytics Engine data about the events and their impact (both monetary and reputational) on the Next, analytics (relational, behavioral, and organization can be captured and used to predictive) are applied to derive concrete predict the amount of savings that can hypotheses (see Figure 4). be realized. Step 3: Geo-locational tagging Figure 3 Rich Profile Data (Location Attributes and Past Events) Twitter RS S Fe ed Improved s Data Google Mash-ups Analytics Response Landing Text Store for Geo-locational Reports & Zone Analytics Engine Times & Cost (ODS) Tagging Dashboards Savings Identification ds Fee Media & Classification S RS Websites32
  • 33. Step 4: Analytics engine Figure 4 Rich Profile Data (Location Attributes and Past Events) Twitter RS S Fe ed Improved s Data Google Mash-ups Analytics Response Landing Text Store for Geo-locational Reports & Zone Analytics Engine Times & Cost (ODS) Tagging Dashboards Savings Identification s eed & Classification SF Media RS Websites n Information about their quarterly and Reporting and Dashboard annual results n User experiences about the vendor onOnce created, scenarios can be displayed to various public or domain forumsusers in a variety of formats and media,customized for various units within the Next, Bank-A applied the geographic codesorganization. Possible formats include: to examine location-wise performance of the vendors.nDashboardsnVisualization charts Using the filters and geo-coding of the data collected, some of the early warningsnOn-demand reporting identified for Vendor-X were:Additionally, data can be integrated with the n stock price A fallingvarious internal systems of an organization,and algorithms can be implemented to n was losing projects to its Vendor-Xgenerate automatic alerts to key individuals, competition in major marketsbased on various triggers. Figure 5 (on thenext page) shows the end-to-end layered view n Regulatory bodies slapping fines onof the platform. Vendor-X for non-complianceComing full-circle to the hypothetical n attrition rates at senior Highscenario highlighted earlier in the article, we management levelscan see how Bank-A identified the risk of n user feedback in many forums Very badVendor-X proactively and adapted to it quickly. across the geographical locationsFirst, Bank-A created filters on the various Bank-A also modeled specific comparisonvendors they work with (Vendor-X being one scenarios for Vendor-X and its competitors.of them). The filtering criteria were composed Based on the historical trend compared to itsof the following: competition, they were able to tag Vendor-X asn stock price fluctuationVendor high risk.nNews about the vendor in various media Using what-if scenario modeling on theand social networking sites robust analytics engine, they were able to 33
  • 34. End to end view of the analytics platform Figure 5 Presentation Layer Web Services Business Users Reporting Layer Applications Internal Analytics Layer Machine Text Classifier Learning Data Integration and Management Layer Operational Data Store Standardization Data Quality & Modeling Workbench Language Domain Lexicon Lexicon Business Domain Users Data Extraction Experts Unstructured Data Sources Twitter MediaSites quickly assess the impact across all the Since Bank-A was looking at the data business functions and locations, and on a real-time basis, they were able to quantify the risk in terms of projects, identify emerging trends and work with resources and dollars. their internal teams to de-risk Vendor-X quickly and move to alternate plans before Vendor-X collapsed as a company.34
  • 35. Conclusion ReferencesMost large organizations have risk functions "Realtime Twitter". their respective units, looking at risk and 8/2/2010. mitigation planning. Most of these 2010/08/02/realtime-twitter/activities are performed manually and are Gigatweeter. Unfortunately, thedecision-making process is limited by Solis, Brian. "Brian Solis Introduces thethe information made available to them Conversation Prism".through restricted traditional channels. Marcomprofessional.comMoving to more automated, fact-based, End-to-end layered view of the platformreal-time monitoring that utilizes the referenced from Infosys HIMI Advancedpower of digital media empowers an Analytics Platformorganization to plan for resilience well inadvance. This further helps shift theorganizations focus to increasing thequality of services offered to customers. 35
  • 36. Analytics in Financial Services05Why Real-Time Risk Brad Jolson Senior Director of Cecilia Mao Director of Product James Patterson Principal Consultant, Product Management, Management, FICODecisions Require FICO FICOTransaction AnalyticsThe double shock of recession-spurred delinquencies and new regulatory inroads on profitis making creditors acutely aware of the need to make sharper risk distinctions amongcustomers. In both account management and collections, they need precise insights to guidemore targeted, timely actions. Every bank with transacting accounts has the potential toachieve this higher level of risk precision, using credit card and debit card account transactiondata. This article demonstrates how combining transaction scores with traditional behaviorscores and credit bureau risk scores increases accuracy of risk predictions. It also discusses theadvantages of deploying transaction analytics in real-time mode, to accelerate awareness ofdeveloping risk and to enable early intervention to mitigate losses. intervention with clients at the point-of-sale, Introduction when early signs of improving or deteriorating performance occur. In addition, many banksWhile top banks have recently increased their use models that capture only a portion ofuse of transaction data, even industry leaders the insights that could be extracted from theirhave yet to realize its full potential for real-time transaction data.detection of changing risk. The means to make sharper risk distinctionsTransaction analytics detect changes in risk in real-time are available today. Banks can takeas they occur, and generate fresh scores with advantage of better modeling techniques, aseach transaction. Yet, most banks do not have well as software that enables deployed modelsthe technology to store and process every and capture of transaction history, to delivertransaction in real-time, and they process the accurate scores in the milliseconds it takesscores in batch mode, usually once a month. for authorization of a credit card purchase orThis batch processing prevents promotion or an ATM withdrawal.
  • 37. Increasing Decision patterns for all consumers (dark bars), Accuracy and Speed perhaps as a result of economic stress. However, accounts that were seriously The combined use of traditional behavior delinquent by the 2008–2009 period showed scores and the FICO® Score to achieve a more much more pronounced changes (orange accurate assessment of customer risk has been bars) during this period. These customers standard industry practice for many years. have changed their spending as the economic The predictive models that generate these cycle has changed, including higher levels of scores analyze different data sources and, cash advances, and reduced spending in travel consequently, provide different risk and home improvement. perspectives. It is the combination of spending behaviors Transaction analytics lift performance even that creates a pattern indicative of rising or further because they model an additional rich falling risk, and sometimes it is the absence source of data and provide an additional risk of activity that is most indicative. For perspective. Authorizations of credit and example, as consumers become more risky, debit account transactions contain abundant their spending shifts away from categories detail—the what, when and how much of such as home improvement and retail. Single customer spending—from which patterns event transactional triggers (e.g., the third cash advance in a month) miss these clues to indicative of risk can be drawn. changing risk. To capture these changes Some of these patterns are evident in Figure 1, related to the decrease or absence of activity which shows percentage changes in credit requires a technology platform that captures card authorization amounts from a six- transaction history and trends. month period in 2007–2008 to the same Transaction analytics—models that capture period a year later as the recession set in. the most indicative characteristics and In this chart, we see changing spending relationships in transaction data—detect such Revealing risk in customer spending patterns Figure 138
  • 38. spending patterns. As a result, they are able to is that the kinds of risk patterns shown inseparate accounts that appear the same in Figures 1 and 2 are detected as they develop.cycle-end activity summaries but actually Indeed, the best models pick up patternshave different levels of risk. indicative of changing risk within a handful of transactions. Because a fresh risk score canThe best models are very sophisticated and be generated with every transaction, it reflectsvery specific. They analyze characteristics that spending patterns drawn from behaviorgo beyond whether the consumer bought taking place a day before, or even minutesgroceries or gasoline to, for example, where before the current transaction.the consumer bought groceries and what timeof day or night the consumer bought Figure 3 shows the potential advantage ofgasoline. They capture additional dimensions real-time transaction scoring. Here we see oneof time and space, such as velocity of cash accounts transaction scores, generated withadvances and changes in elapsed time for each new transaction, as well as the behaviorvarious types of purchases. score at the same point-in-time. The falling transaction score shows a significant changeAs shown in Figure 2, transaction patterns are in risk, which would otherwise not be pickedcustomer-specific; therefore, analytic models up until cycle-end when a new behavior scoremust also be able to differentiate between is generated. Real-time transaction scoring,patterns that are unusual for one customer, but however, gives the bank the opportunity tonot for another. Making an ATM withdrawal at stop transactions as early as 11/ 15 as shown11:30 pm may be a departure from one below—23 days before the end of the cycle.customers normal spending pattern, and itmay indicate a changing risk profile, while it Research on Combinedmay be completely ordinary for another Score Performancecustomer, who works on a night shift.The most important point to understand New FICO research demonstrates the addedabout todays advanced transaction analytics value of transaction analytics for reducing losses. Customer spending patterns may reveal risk Figure 2 39
  • 39. Timely alerts to changing risk Figure 3 In the first two studies, combining a Sharper Credit Line transaction score with a behavior score and Decrease Decisions FICO® Score increases risk differentiation, enabling finer, more accurate segmentation In the first data study, comprising of more in credit line management and collections. than 10 million North American credit card In the third study, a transaction score used accounts, FICO examined the predictive lift with a behavior score accelerates creditor from the combined use of behavior scores, the awareness of accounts showing signs of FICO® Score and transaction scores in financial distress and enables timely credit line decrease decisions. The results interventions to contain losses. demonstrate significant improvement in risk discrimination and, in particular, markedly Increasing segmentation granularity for credit line decrease Figure 4 by adding scores40
  • 40. greater risk purity of end strategy segments into transaction score quintiles. This tri-scoreacross current and mildly delinquent approach expands the number and precisionaccounts. of risk-separated segments, with bad rates now ranging from 2.33% bad rate to 35.90%.For simplicity and brevity, the chartsbelow show the impact of combined scoring This fine-grained segmentation enables theon just one slice of the accounts: those portfolio manager to assign credit linewith behavior scores in the lowest- decreases with much greater precision. Figurescoring 10% (decile 1). This high-risk segment 5 shows a strategy diagram with segmentsrepresents about 6% of the entire account receiving line decreases in orange. Accountspopulation. receiving no decrease are in green.As shown in Figure 4, the aggregate bad rate In this diagram, we can see that:for the accounts in this behavior score decileis 13.53%, where “bad” is defined as accounts nthe behavior score and FICO® Whenwith a maximum delinquency of 3+ cycles, Score are used together, accounts fallingbankrupt or charged-off in the subsequent in the lowest 60% of FICO® Scoressix months. (deciles 1–3) receive the decrease. These ® are represented by the three orange boxesAdding the FICO Score refines the view of ® in the FICO Score row of the diagram.risk, enabling more granular segmentation.The five resulting segments, based on 20% n transaction score is used with When the ®FICO Score divisions (quintiles), have a bad the behavior score and FICO® Score, therate ranging from 5% to nearly 27%. greater precision in risk separation enables the strategy to be fine-tuned. WithAdding the transaction score furtherrefines the view of risk, increasing this added precision:segmentation granularity even more. Each of About 12% of accounts from the ?the FICO® Score quintiles is broken apart decrease population no longer receive Strategy for credit line decrease Figure 5 41
  • 41. the line decrease. These “swap outs” are of transaction scoring on the segmentation represented by the green boxes at the of delinquent accounts, using the same 10 bottom of the second and third million North American credit card accounts columns from the left (under the as the previous study. For simplicity and ® orange FICO Score boxes labeled brevity, the following charts show the impact 15.52% and 11.50%). of combined scoring on just one slice of the accounts—one-cycle delinquent accounts ? of accounts not previously About 8% in the middle-scoring 10% (decile 5) of targeted for line decreases now receive behavior scores. The study demonstrated them. These “swap ins” are represented particular benefit in moderate score ranges by the two orange boxes at the top of such as this—those traditional “gray areas” the fourth and fifth columns (under of risk, where portfolio managers may have the blue FICO® Score boxes labeled the least confidence in applying treatment 8.29% and 4.97%). to accounts. Better segmentation of accounts for credit limit decreases and other loss mitigation As shown in Figure 7, the aggregate bad rate treatments can have a significant financial for the accounts in this behavior score decile impact, as shown in Figure 6. is 23%—where “bad” is defined as accounts with a maximum delinquency of 3+ cycles, bankrupt or charged-off in the subsequent Sharper Collection Treatment Decisions six months. ® Adding the FICO Score refines the view of FICO also performed a study on the impact risk, enabling more granular segmentation. Estimating the benefit of sharper credit line decrease decisions Figure 6 Sharper segmentation enables you to prevent losses by avoiding balance build on riskier accounts. Figure 6 shows results in three estimated lift scenarios. These estimates show only one side of the benefit to be gained from more accurate credit limit decrease decisions—sharper segmentation also improves profitability.42
  • 42. Increasing segmentation granularity in one-cycle Figure 7 delinquent accountsThe five resulting segments, based on 20% the view of risk, increasing segmentationFICO® Score divisions (quintiles), have a bad ® granularity. Each of the FICO Scorerate ranging from 13% to nearly 41%. quintiles is broken apart into transactionAdding the transaction score further refines score quintiles. This tri-score approach Strategy for prioritization of collections treatments Figure 8 43
  • 43. expands the number and precision of risk- and transaction score are all used, the separated segments, with bad rates now greater precision in risk separation ranging from 5% to 51% bad rate. enables the strategy to be fine-tuned. Now only 8% of accounts receive This sharper segmentation enables collection managers to make better decisions about high priority treatment—enabling where to focus resources. Figure 8 shows managers to concentrate their most various tiers of collection effort, with the experienced and skillful collectors highest one-cycle collection priority in red where the risk is greatest. Another and the lowest priority in blue. 8% of accounts are identified for low-priority treatment, which may The diagram and Distribution of Accounts consist of allowing them to self-cure, table below it show that: and thereby avoiding spending money n the behavior score and FICO® When to annoy valuable customers. In Score are used together, the increased between, accounts are also better separation results in 20% of accounts separated by risk, enabling managers receiving high-priority collection to more finely target, test and evaluate treatment, 40% receiving moderate- appropriate treatments. priority treatment and another 40% Better segmentation of accounts for receiving low-priority treatment. collection treatment can have a significant nthe behavior score, FICO® Score When financial impact, as shown in Figure 9. Estimating the benefit of sharper collections decisions Figure 9 Sharper segmentation enables better collection prioritization, which results in lower roll rates for early-stage delinquencies. Figure 9 shows results from transaction scoring for a standard portfolio of 1 million accounts. By reducing the 1-to-3-cycle roll rate, transaction scoring reduces annual charge-offs by US $1.5 million.44
  • 44. Faster Intervention to that only one type of intervention—terminating Prevent Losses account utilization—is taken. Cycle-based behavior scores and real-timeIn a third study with a top UK card transaction scores were independentlyissuer, FICO sought to determine if the analyzed to determine four cutoff pointsaddition of transaction scoring to behavior each, equivalent to cumulative bad rates ofscoring would accelerate awareness of 5%, 10%, 15% and 20%. The interventionrising risk in distressed accounts and, if is triggered whenever an account reachesso, how much loss mitigation value it would either the behavior score cutoff or theprovide. The previous two studies look at transaction score cutoff within these bad-the benefits from improved accuracy alone, rate categories.not taking into consideration the benefitsof inter-cycle scores and real-time decisions The transaction scores triggered accountthat would add to the overall value of intervention an average of five days aheadtransaction scores. Timeliness benefits from of the behavior scores. Although thatinter-cycle and real-time decisions are hard difference seems insignificant, theto estimate, and this study provides one angle opportunity for loss mitigation was not.of benefit from the credit management view. When the transaction scores triggered account intervention first (27% of allThe study was conducted on a random accounts and 25% of the bad accounts), thesample of 850,000 accounts. The overall changing behavior patterns they detectedportfolio bad rate (3+ cycles delinquent, were indicative of substantially rising risk.charge-off or bankrupt) was 2.6%, compared In those cases, the transaction scoresto 2.7% bad rate average for all the UK for identified preventable balance buildthe same period. averaging US $516 per bad account itFor the purposes of clear measurement in triggered. When the behavior scores triggeredthis study, a simplified scenario was created: account intervention first (21% of allthe study assumes that score cutoffs are accounts and 18% of the bad accounts),the only trigger for intervention used, and the preventable balance build was much less, Early alerts to significant risk Figure 10 45
  • 45. Added value of timeliness in risk scoring Figure 11 only US $28 per bad account. (Note: the and sharper real-time decisions on how to benefit analysis does not include the treat customers, from account management remaining 52% of the population, where through collections. both scores triggered account intervention Most banks, of course, are currently at the same time.) interested in transaction analytics primarily for improving control of risk. Conclusion As the economy recovers, however, the opportunities that can be gained through For banks with transacting accounts, fine-grained separation of customers by data from transactions is the most consistent other behavioral dimensions will garner and frequently updated source of more attention. Precise decisions—driven information on their customers. Under by real-time detection of changes in not todays economic and regulatory pressures, only risk, but also in customer needs and it is essential for uncovering the additional lifestyle—will be essential for extending insights and faster detection of change the right credit at the right time to maximize that enable more accurate risk predictions profit in the recovering economy.46
  • 46. Analytics in Financial Services06Ten Questions to Ask of Lisa Kart Director, Karthik Sethuraman Senior Manager, Analytic Product Analytics,Your Optimization Solution Management, FICO FICOWith so many optimization solutions in the market today, how do banking institutionsevaluate which one will work best for their business? This article defines the criteria needed toevaluate alternative solutions and make this determination—from assessing data sensitivities,stress-testing, and leveraging other analytics assets to validating “optimal” points, handlingbusiness trade-offs and deployment. Banks that follow these guidelines, successfullycombining sound methodology, deep domain expertise and the right software, typically see5%–20% profit improvement. 1. How does the solution incorporate the Introduction sensitivities of consumers to various actionsThe adage “choice is the enemy of decision” In order to optimize the actions that you take oncertainly holds true for optimization solutions. your customers and prospects, you first need toAs banks move towards optimizing customer understand the reactions of each consumer todecisions, the many choices available make it possible treatments you can take. Optimizingdifficult to know what works best. based on incorrect assumptions about these “action-effect” dynamics will render theWhile some optimization solutions are optimization ineffective (at best) or grosslypurely software-based “solvers”, others offer incorrect (at worst).more, such as modeling services ordomain expertise within a given decision Incorrect models of consumer behavior— whetherarea. Moreover, choice is not the only on the dimension of response to a product offer,challenge. Historically, many solutions risk from a line increase, or revenue from a pricingare highly academic, and not built to meet action—lead to incorrect optimization results. Inreal-world requirements. other words, “garbage-in equals garbage-out”.So, what are the right criteria to evaluate an Consider Figure 1, which shows three curvesoptimization solution? representing the relationship between expected profit and different levels of loan price for aStart by asking the following 10 questions, particular consumer.designed to help you understand therequirements, and avoid the pitfalls, of decision The orange curve represents the “true” relationshipoptimization. between profit and different levels of loan price.
  • 47. Incorrect optimization results have significant impact on profit Figure 1 Note that a consumers price sensitivity is are influenced by multiple factors. The typically not known a priori, for several modeling techniques required to accurately reasons—for instance, you may not have tried understand consumers sensitivities are very every possible combination of price on this different from traditional scorecard consumer, actions other than price are taking approaches. Ensure that your vendor is fluent place simultaneously, or the environment is in decision modeling techniques, can validate changing. This true relationship becomes its assumptions and approach, and can deal apparent only after the actions are taken, with limitations in data. and balances the likelihood of taking up the 2. Does the optimization solution address loan with the potential revenue and loss. data limitations The red line represents a “good” estimate, Getting the action-effect relationship right since it estimates the true relationship closely. depends on both data and business expertise. The green line shows an incorrect estimate Data will never be perfect, and any vendor that is a bit off, perhaps due to data biases. who claims you must invest 6–12 months in Without constraints, the optimal action is running experimental designs to gather determined by the top point on the curve. If a perfect data is mistaken. Leveraging your lender took action based on the incorrect existing data is important, and can tell you estimate, it would offer this consumer 9.49%, a lot about your customers and their response while an action based on the good estimate dynamics. would offer 10.49%. Looking at these points While utilizing your data assets is important, on the true profit curve, the good estimate ensure that the solution does not blindly rely would increase profit by about US $4 per on historical data, but instead has procedures account. Multiplied over hundreds of in place to tease out casual relationships thousands or millions of accounts, the effect despite biases in the data. The ability to of the error on profit would be significant. identify and address these holes and/or biases Action-effect models are typically more in data is an important component of your complex than the examples above, since they solution, since these can lead to incorrect48
  • 48. assumptions about consumer behavior, as you are confident all relevant customer data isillustrated in Figure 1. utilized to make the best decision. Your data elements and models can be used as inputs orIn the rare cases where no data is available, decision keys, and in many cases, can be usedthere are still ways to move forward directly in the underlying action-effecttoday, including: 1) Developing expert models themselves.models by encoding business expertiseinto the assumptions and relationships, and 4. Does the solution provide insight2) Performing smart, limited testing to into your key business trade-offs toquickly gather the data to inform your facilitate the selection of an optimalmodels. In cases where you need to rely operating pointheavily on expertise, more stress-testing Your business will need to optimize decisionsmay be required, which will be discussed subject to many constraints. These rangefurther in question 7. Choose an from policy constraints, such as who you lendoptimization solution that explicitly to or what offers certain customers are eligibleidentifies and addresses data limitations, and for, to portfolio-level goals, such as “reducecan incorporate business expertise. losses by 5%” or “ensure marketing budget is3. Can you leverage your existing analytic less than US $10 million”.assets within the solution While some optimization solutions focus onYou have invested in gathering data, finding the “one” solution to a problem youbuilding predictive models and perhaps even specify, it is always important to understandbuilding action-effect models to drive your the impact business constraints have on youroptimization solution. The optimization bottom line and your strategy. Exploringframework that you choose should allow these trade-offs helps to properly set thoseyou to incorporate these models in a way that constraints. Efficient frontier adds critical insight into key business trade-offs Figure 2 49
  • 49. A key tool for quantifying these trade-offs today. For instance, “reduce attrition by 10%” is an “efficient frontier”, as seen in Figure 2. will only be meaningful if you tie it to the For example, to increase both profitability current outcome. and market share, you may look at the rate at In some cases, the “business as usual” baseline which new loans are “funded” (approved and is best represented directly from the data. opened). These historical actions can be passed directly If today you are operating at point A, you can through the simulation. This option tends increase both profit and market share by to work best when multiple strategies moving to the efficient frontier (e.g., move up (e.g., champion/challenger) are being used. and to the right to point C). In other cases, it is helpful to represent your If you are operating at point B, the trade-off baseline strategy as a set of rules or a decision between business goals really comes into play. tree, particularly when this strategy is not You may decide to maintain your current represented in the data. market share and increase profit (by moving up A good optimization solution will provide to point D). Or you may decide to maintain or both options for simulating your current even sacrifice profit to gain additional market strategy or even a strategy you are considering. share (by moving to points E or F). 6. Does the solution provide a mechanism Similar trade-offs can be made between profit for business users to verify that the and loss, volume, attrition, exposure or any optimized solution is valid other business metrics of interest. Optimization results are only as good as the A trusted advisor with relevant business assumptions and inputs on which they were expertise can help you identify the optimal built. As such, it is imperative to evaluate the operating point. Make sure your optimization optimization results from both a technical vendor can provide someone you can work and business point of view. with to identify options and make a choice on A mechanism to investigate optimization strategies to deploy. results allows the business user to impute 5. Does the solution allow you to establish business judgment into a process that is and evaluate a baseline of “business as otherwise a pure mathematical exercise. usual” or what you are doing today Let us look at an example for offering a new There are many reasons why you should be line of credit. Figure 3 compares customer able to measure your current strategies in the profiles generated by an optimization context of the optimization solution. First solution: Customers who are rejected, and foremost, it provides a Return on customers who are accepted with a low line Investment (ROI) estimate of the optimized of credit, and customers who are accepted solution. It is important for this estimate of with a high line of credit. The goal is to see lift to be made within the optimization whether the more favorable treatments will solution to ensure an “apples-to-apples” generate the profit improvements that the comparison. It also provides a sanity check solution projects. that your decision model is able to project the Business experience suggests that Example 1 results of the historical strategy. may be too optimistic. For instance, the Another reason is to tie the various outcomes “accept with high line of credit” group is still ® and constraints to what you are experiencing fairly risky (judging by FICO Score and debt-50
  • 50. to-income ratio) and their need for credit is business knowledge? In other words, is therelatively low (37% utilization). It seems optimization driven by unreasonableunlikely that the solution would generate modeling assumptions?many gains from this treatment. To help the business user validate theIn Example 2, the solution identifies less risky reasonableness of the approach, theaccounts with higher credit needs for the optimization solution should have reportingmore favorable treatment. Therefore, it is functionality that provides visibility andmore likely to generate profit. hands-on review. “Swap set” reports can also be useful to compare different optimizedThe ultimate question is: Can the business strategies, or an optimized strategy to whatuser justify the treatments based on his/ her you are doing today. Optimization results must make business sense Figure 3 51
  • 51. Ideally, the business user could create his/ her 8. Can the optimization framework be own diagnostic reports in the tool and drill easily adopted for new decision areas down, where necessary, to get comfortable across the enterprise with the optimization. Some optimization frameworks require 7. Can the optimization solution help you the user to formulate the mathematical better prepare for potential changes in expressions corresponding to the optimization your environment problem. While applications made for a mathematically oriented “super user” Despite investing in good data, software allow flexibility, the trade-offs are that it and expertise, unexpected events can happen can be more difficult to use, further from to change the business world. The economy the business problem and often very changes, the competitive environment error-prone. Using this methodology, it is changes, the way consumers react changes. much harder to validate mathematical This means the outcomes expected from expressions than to validate simple business the strategy—profit, revenue, loss, market logic. In addition, it is much more difficult share, to name a few—will also change. for the business user to understand and use Ideally, the optimization solution captures the framework, and may require specialized the interaction between various macro- training or additional resources. economic conditions and underlying Conversely, if the framework is too model estimates—at a minimum, for the risk restrictive, you will not be able to leverage models. When done, you will be able to it in other applications and decision areas, evaluate the impact of various exogenous which is a common goal for many building effects. This will help in choosing strategies an optimization practice. For example, you that will fare well under a variety of might want to optimize your marketing market conditions. This is not only decisions after you tackle your line increase important from a regulatory perspective decisions. (as required by the Basel II Accord), but it Therefore, it is critical to balance flexibility also makes good business sense. within skill set requirements when choosing Thus, an optimization solution must an optimization solution, and have a tool provide a facility to perform “what-if” that both meets analytic requirements and scenario analysis that goes beyond simply communicates to the business user. This will changing your constraints. It should be allow your organization to scale your able to simulate the impact of changes, optimization efforts from a common such as external factors (like the economy) methodology and platform, and develop a or consumer behavior (like price sensitivities repeatable process that can grow as your or risk levels). business grows. The best solutions will allow you to: 1) 9. Can the optimization solution be readily Automate stress-testing to perform many deployed in your existing decision-making scenario investigations at once, 2) Assume platform that changes may apply differently to After choosing the optimal scenario to different segments of your population, and 3) be deployed for your portfolio, the next Provide a range of insightful diagnostics and step is to bring those optimal actions or support to inform your choice of optimal rules to your current decision execution strategy, in light of what the future may hold. platform. In financial services and elsewhere,52
  • 52. this is most often in the form of a strategy tree Many optimization software vendors try tothat can assign treatments to current and circumvent the problem by optimizing at thefuture customers. However, in certain segment level or using an approximate searchcircumstances, it may mean calling the algorithm—relaxation of the optimizationoptimization from within an application to problem, as it is known in academia. The bestoptimize real-time. optimization algorithms can exploit the structure of a particular decision problem,If you are deploying to a rules management while still solving for the true, rather thansystem, it is beneficial to have software that approximate, solution.can export these rules in a format that yourdecision-making platform can accept, such as In financial services, it is reasonable to assume acode. This step eliminates the errors that problem size of millions of consumers oroften happen when trying to manually enter accounts, with dozens of business constraintsthe optimized strategy, which can be much (in practice, 2–5 portfolio constraints are mostmore complex than a typical strategy. typically applied), and dozens to hundreds of possible treatments. Depending on yourThe key to deployment is flexibility and computing power and number of constraints,the ability to translate optimal actions to be this size problem should take a few minutes to aused by your system. few hours to solve.10. Does the optimization solution have Finally, choose a vendor that continuallysophisticated optimization software invests in its optimization algorithm todesigned to solve business problems of improve performance and servicesOptimizing decisions on millions of customers High-value Optimizationwith multiple business constraints—as is donein financial services—requires an optimization Evaluating an optimization solution usingalgorithm that can find optimal solutions the guidelines above can help you avoidefficiently. Your optimization algorithm must potential pitfalls and ensure that yoursolve for the best (i.e., true optimum) solution solution delivers true business value. Theseand exploit the structure of your decision questions were crafted using best practicesproblem to do so efficiently. Questions to ask from FICOs more than 100 optimizationyour vendor include: projects across different countries, portfoliosn algorithm solving for theIs your and applications. Recent client results include:true customer-level optimum, or n 12% profit improvement in credit lineapproximating it? management, while controlling losses.nHow quickly can you solve problems with n improvement among new loan 45% profitmillions of customers and dozens of applicants by improving pricing andconstraints? loan amount decisions, while meetingn do you invest in Research andHow much regulatory requirements.Development, and improvement of the n increased profit per eligible US $10software and algorithms? customer through better targeting ofnhave any performance benchmarksDo you marketing offers, within the samefor your optimization technology solving marketing budget and risk profile.challenging optimization problems? 53
  • 53. Banks that are most successful with decision optimization combine sound methodology, deep domain expertise, and software designed to address the mathematics and business implications of optimization. Depending on the constraints imposed, these companies typically see 5%–20% profit improvement.54
  • 54. Analytics in Financial Services07Practical Challenges of Lisa M. Kart Director, Mac Belniak Director, Analytic Product Analytic Pre Sales,Portfolio Optimization Management, FICO FICOThe word “optimization” is often used informally to describe any technology that improvesbusiness results. In reality, optimization is a mathematical methodology used to makedecisions for allocating finite resources to achieve an overall objective, subject to theconstraints imposed by the environment. While the theory of optimization is well studied byacademics in fields ranging from operations research to computer science, there are manypractical challenges when applying optimization to develop superior decision strategies. Thisarticle examines these challenges and demonstrates how optimization can be applied to solvereal-world business problems, using an example from mortgage lending. Although optimization can be applied broadly Introduction across a range of industries, this article uses a mortgage loan remediation case study toTraditionally, organizations have applied methods demonstrate key points. In this example, thesuch as champion/ challenger to evolve decision servicers goal is to maximize life-time Netstrategies. A decision maker works to develop a Present Value (NPV) from the portfolio of loans,candidate strategy (challenger) that has the while offering loan modifications andpotential to surpass the status quo (champion). minimizing re-default risk. For each mortgageAnalytics and intuition are used to assign actions loan borrower, the servicer needs to take lossthat may yield better results, while still adhering to prevention measures and protect the NPV of theconstraints. The challenger is then tested against portfolio. The servicer can consider the merits ofthe current decision strategy to determine which is refinancing the loan, modifying the loan terms,better. The decision strategy is improved iteratively pursuing a short-sale, or foreclosing on theby developing and testing new challengers. Much property. Proactively presenting an alternativeiteration is required to find the best possible loan offer (i.e., changes to interest rate, term,strategy, making the process of improving a principal amount, or any combination) canstrategy time-consuming. influence the behavior of the borrower, in termsAn approach based on optimization provides of ability to make the monthly mortgagea path to finding the best possible challenger payments, reducing the likelihood of walkingstrategy much faster. This article provides away, and ultimately protecting the NPV of theinsights on the common issues encountered loan. However, deciding among the alternativeswhen applying optimization to complex requires balancing the needs of the homeownerdecision scenarios. and the profitability needs of the investor and
  • 55. servicer, which is challenging. Focusing on historical data, facts, scores, predictions, the servicers goal to maximize NPV, the forecasts, and uncertainties. This information challenge is in making decisions across the must be marshaled into an optimization entire portfolio in anticipation of model that achieves a specified goal and delinquencies, losses, recidivism and adheres to constraints while considering revenue—while managing uncertainty about uncertainties. Formulating a well-structured the performance of the individual loan and optimization problem from this wealth of the portfolio as a whole. business knowledge is quite challenging, and often requires skill and expertise in Optimization in Action operations research and computer programming. It is useful to introduce a framework for this task from the field of This section explores the practical considerations decision analysis—namely, a decision model. in applying optimization, from problem formulation through deployment. Decision models incorporate all of the information that is important for Formulating an Optimization Model formulating and executing optimization. Optimization balances a multitude of A decision model maps the relationship entities, such as goals, policies, constraints, between information, such as historical data Influence diagram for mortgage loan remediation Figure 1 In this mortgage loan modification example, the decision model identifies how each borrower (with a FICO® Score, income, current loan amount, etc.) is likely to respond (probability of walk-away or expected loss given foreclosure) to all of a servicers possible actions (change in interest rate, term, etc.), and what the result will be (NPV). Based on this decision model, it can analytically be determined which action to take for each borrower in order to optimize NPV, given business constraints on the portfolio, such as acceptable losses or cash flow targets.56
  • 56. and facts, and prediction. It relates possible example, decrease the interest rate inactions to their effects and likely outcomes, increments of X basis points. When there areand outcomes to business goals and multiple decisions, it is practical to introduceconstraints. In this sense, a decision model is the concept of treatments. Treatments specifya unified model of the decision situation combinations of these decisions—forfacing the business. example, offering a 2% reduction in interest rate and extending the term by five years.Decision models abstract the complexity offormulating optimization problems to make Constraintsadvanced optimization accessible to business A practical challenge is that mostusers, statisticians and analysts. Decision organizations have policies that limit themodels can also be represented graphically by alternatives that can be considered. Theseinfluence diagrams—the decision models constraints come in two varieties, record-include mathematical formulae for the level/ account level and cross-recordrelationships shown in the influence portfolio/ segment level. Record-leveldiagram (Figure 1). constraints place restrictions one record at aA key challenge in decision modeling is to time. In account management, there is onedescribe the effect of taking alternative record per customer. There are a number ofactions. In financial services, this is often policies that may apply at the accountaccomplished by developing predictive level—for instance, never send the same offermodels called action-effect models—for to a customer two months in a row. Record-example, a model of the effects of modifying level constraints may also be used to specifythe loan on key factors such as the likelihood treatments (combinations of decisions)—forof re-default or foreclosure. At times, some of example, do not offer a loan modification tothe actions may not have been taken before, someone with debt to income <31% or do notand thus, limited data is available to build a offer principal reduction to a borrower with amodel. Experimental design techniques can loan to value < employed to collect data to fill this void. The second type of constraint applies acrossIt is also desirable to have a software multiple records. In loan modificationenvironment that can use all of the data decision problems, there are a handful ofavailable, yet allow a human expert to extend cross-record constraints. Some apply acrossthe capabilities of the model to predict the entire portfolio—for example, reducepreviously untested actions. Without expected portfolio losses by X% or reducecustomized software and a robust portfolio foreclosure rates by Y%. Whenmethodology to incorporate data and cross-record constraints apply to all thebusiness expertise, developing these action- records, they are sometimes called global oreffect models is cumbersome at best. portfolio-level constraints. Other cross-recordMultiple Decisions and Treatments constraints might apply only across specific segments—e.g., losses for the high-riskMost decision situations involve two or more segment must be less than US $Z next year.decisions. For example, in a loan remediationprogram, there are many possible decisions: The values of all the controllable and non-offer a reduction in interest rate, reduce the controllable variables must be known inprincipal amount, extend the term of the loan order to check whether a cross-recordor some combination. Each decision has a constraint has been satisfied. Constraints thathandful of alternatives for each customer—for apply across the entire portfolio, across 57
  • 57. segments or across more than a handful of profit while maintaining acceptance volume records add significant complexity to the of new loans, or maximizing NPV while optimization problem. While business rules allowing only a limited number of management technologies and many modifications suited for the business optimization packages are capable of environment. In practice, problems with enforcing and accounting for record-level multiple objectives can be reformulated to constraints, they are not capable of finding have a single objective. This is done either by the optimal decision strategy in the presence forming a weighted combination of the of cross-record constraints. However, once an different objectives, where the weights represent optimal decision strategy has been developed, their relative importance or utility (Keeney, business rules management software can be 1993); or by representing some of the objectives used to maintain and execute the strategy, as constraints whenever those objectives must allowing it to be put easily into production attain specific values—for example, maximize (see Deploying Optimization). NPV subject to the constraint that cash flow reductions not exceed US $X million. Multiple Objectives Sensitivity Analysis When specifying the objectives for an optimization, there are almost always multiple Framing an optimization problem in terms conflicting goals—for example, maximizing of a single objective and multiple constraints revenue and minimizing risk, maximizing sets the stage for a sensitivity analysis that can Efficient frontier of NPV vs. Accounts modified Figure 2 The choice of the operating point on the efficient frontier can be selected by considering the subjective trade-off between NPV of portfolio and % of portfolio modified. If the goal is to purely maximize NPV for the portfolio, the servicer might choose a higher percentage of accounts for modification. However, if there are some constraints on numbers of accounts that can modified in a month, it may choose a lower point on the curve.57
  • 58. yield valuable insights. By repeating the performance of other customers. The second isoptimization with different loss constraint uncertainty that affects the behavior of thethresholds, the trade-off between profit and entire portfolio simultaneously in the samethe threshold can be evaluated. way. For example, an unexpected economic downturn may impact the ability to makeConsider the trade-off between expected NPV timely payments for a large number ofand percentage of delinquent loans modified. borrowers in the portfolio or segment at theFigure 2 illustrates that the NPV is lower with same time. It is important to account for bothfewer modifications and can be elevated by types of uncertainty when optimizingincreasing the percentage of loans modified. decisions. This helps to ensure that theHowever, there is a point of diminishing decision strategy will be robust.returns, whereby modifying a higherpercentage of loans does not result in an Academic literature offers several approachesimprovement in NPV. to such decision making under uncertainty within the context of constrainedStress-Testing optimization. FICO favors robustSensitivity analysis demonstrates how the optimization techniques (Mulvey, 1995).objective value and optimal strategy change “Stochastic” or robust optimization,across different constraint thresholds. Stress- involving probability analysis, provides antesting shows how a given strategy and its effective means of representing uncertainty,objective are affected by changes in key when it is possible to generate appropriatevariables and uncertainties. Stress-testing can sets of scenarios and scenario used to evaluate how robust the objective This approach seeks solutions that maximizevalue of the optimal strategy is. For example, the expected value of the objective, andassume that there is an optimal loan therefore, provides the best answer across allremediation strategy that is expected to scenarios (Clemen, 1990). Ideally, theincrease the portfolio NPV by US $Y million, scenarios are formed and their probabilitiesbut one would like to know the range or are assigned in harmony with the experiencedistribution of NPV to expect when executing of the decision maker, as well as the patternsthe strategy in an uncertain economy. Stress- that are evident in the available data.testing can compute the objective for a wide There are also different formulations forrange of values given uncertain driving modeling risk neutrality or risk aversion onfactors such as interest rates, yielding a range the part of the organization (Keeney, 1993).of NPV. It may be desirable to choose a The critical point is that many problems arestrategy where the objective is robust in the inherently rich with uncertainty; robustface of a wide range of circumstances. optimization techniques adequately address that uncertainty, while simultaneouslyUncertainty allowing for the inclusion of a long-termAnother practical aspect of many objective function, and both record-level andoptimization problems is the management cross-record constraints.of uncertainty. As with constraints, there Deploying Optimization—Reaping theare two basic types of uncertainty. The first Benefitsis uncertainty at the record level. In accountmanagement, there is uncertainty surrounding The final practical aspect of applyingthe performance of an individual customer. optimization is deploying the optimizationThis uncertainty is independent of the in the production environment. Often, it is 59
  • 59. sufficient to deploy the results of an optimizations can be run on production data optimization based on the data used in to assign actions. Each run can then developing the optimization model. One consider all of the records, including any can either deploy a database that contains new additions. actions assigned to each record or develop a decision tree that generalizes the optimal Conclusion strategy. In effect, these tables or trees are deployable action plans that FICO calls While this article included a mortgage loan strategies. Exporting a database that contains remediation case study, the benefits of actions assigned to each record ensures that optimization are not limited to mortgage all cross-record constraints are satisfied by the lending. FICO clients in other areas of business strategy, but also requires that all banking have also seen improvements records be considered in the optimization. across the customer lifecycle. By optimizing Results, at times, need to be generalized credit line strategies, for example, one card so that they can be applied to records that issuer has seen US $12.36 increased profit have not been considered to date. For per active account in the first 12 months. example, in account management, a customer By optimizing the price and amount of may be acquired after the optimization. new loan originations, one lender saw In this case, the challenge is to develop a 45% profit lift over the life of the decision tree or a set of business rules that loan. Ultimately, optimization—with the encodes the optimal actions, while still practical considerations addressed in this adhering to the cross-record constraints. article—enables organizations to improve A set of rules or a decision tree guided by performance by more quickly identifying optimal actions is a way to generalize the the best offer for each customer, while optimization action assignments for balancing objectives under existing business taking a consistent set of actions on an constraints. ongoing basis. Decision trees are also extremely useful in References visualizing and interpreting the action plan. They provide unprecedented visibility into 1. Keeney, R. and H. Raiffa, “Decisions exactly how each decision is reached, as well as with Multiple Objectives”, Cambridge an understanding of the key drivers of value. University Press, Cambridge UK, 1993. For practical applications, it is extremely 2. Clemen, R., “Making Hard Decisions: important to be able to deploy optimization An Introduction to Decision Analysis”, results directly to applications using Duxbury Press, Belmont, California, business rules products or decision systems, 1990. such as a customer management solution. 3. Mulvey, J. M., R. J. Vanderbei, and If many new customers are added to the S. A. Zenios, “Robust Optimization of portfolio on a regular basis, and the data used Large-Scale Systems”, Operations to create the initial strategy no longer Research, 43 (1995), 264-281. represents the profile of current customers, the optimization model can be deployed into the production environment so that60
  • 60. Analytics in Financial Services08 Yamini Aparna Kona Balwant C. SurtiAnalytics in Cross Selling – Senior Consultant, Industry Principal and Infosys Technologies Head-Solutions ArchitectureA Retail Banking Perspective Limited and Design Group, Finacle Solutions Consulting Practice, Infosys Technologies LimitedThe case for cross-selling to the existing customers of a bank is an easy one—the difficultpart is executing it. Today, there are several different techniques for cross-selling effectively.The common thread that runs across them is data and analytics. Predictive analytics basedon various models have created offers that are just right, just in time. Data mining andanalytics have helped in discovering trends and populating models that are the backboneof predictive analytics. Value analytics is another approach to cross-selling that is available.The call center, the branch, the web—every distribution/ service channel—all leverageanalytics in some way to cater to the entire gamut of customer needs—not just what thecustomer seeks. This article analyzes the different ways in which cross-selling workswith analytics, its intrinsic challenges, and the emerging trends in the analytics field. clients becomes increasingly difficult and Why Cross-Selling is Imperative expensive in a highly commoditized industry, selling more products to existing customersThe experience of many financial institutions makes great business sense for a bank. It is anshows that the cost of selling an additional excellent way to increase revenues and indirectlyproduct to a current customer is one-fifth improve customer retention, because customersthe cost of selling the same product to a with more products tend to be more customer. This explains why cross- Customer attrition rates are inversely proportionalselling, i.e., selling a bundle of products and to the number of products held—the more productsservices to the client (usually an existing one), you sell to the customer, the lesser is the chance ofis being increasingly considered the cornerstone the customer leaving you. As a result, movingof the retail financial industry. from a silo-product mentality to a consultativeAs other sources of organic growth (for example, selling approach has resulted in a proliferation ofloan demand) have slowed, and adding new cross-sell initiatives in the banking segment.
  • 61. effective in the hands of a skilled advisor Approaches to Cross-Selling who can extract portfolio-related information from a client. This approach Cross-selling is selling additional products also has the advantage of revaluing the to existing customers or prospects. It may portfolio at periodic intervals and happen along with the initial sale or after coming up with other opportunities for the initial sale is made. Often, the customer cross-selling. may not explicitly mention specific needs 4. Predictive Analytics-based Approach: or be aware that the bank offers products This refers to a set of approaches where a that meet their needs—cross-selling taps into model (or a set of models) characterizes this unmet potential using a variety of customer buying behavior for financial techniques: products. Past customer data is used to 1. Person-based Approach: This is based build, refine and modify predictive on either the skill of the Customer models. These models are used to predict Service Representative (CSR) or through future customer buying—information a structured question-based approach. In used to generate customer offers. either case, the emphasis here is to elicit In many circumstances, current or recent the need through customer interaction. transactions are used as trigger points in Often, the skill of the CSR is the deciding the system, and very often, the current factor of success, and little or no use of customer interaction is used as the means analytics is made. to deliver the offer. Trigger-based models 2. Rules-based Approach: The system can range from simple to sophisticated. defines a set of rules and uses the Advanced versions can analyze a current information collected from the customer online transaction and couple it with past to arrive at a cross-selling offer. Some data to present relevant offers. Offline analysis of the customer data is made. For offers are also often analyzed to come up example, while processing a loan with the best channel for delivery of the application, enough information is offer (for example, by mail, through a available to decide whether the prospect call, etc.) and some offers may be made qualifies for a credit card as well. using a combination of channels used in an orchestrated manner to get the 3. Value-based Approach: This follows a customer hooked (for example, a teaser portfolio approach to the customers mail, with a click to a website or a phone assets and liabilities with the bank. Here, number to call or meet a particular a customer is given a scenario with one branch officer). The success or failure of product that he or she has asked for. an offer is also an input to the model to Then, based on other information improve future success rate. obtained from the customer, alternate scenarios are offered. Certain value 5. Social Networking-based Approaches: metrics (for example, net assets, These are not yet prevalent in retail installments per month, average rate of banking, but here again, a persons social interest paid, etc.) under multiple networks, likes, dislikes, preferences, scenarios with additional products are recommendations from network friends, presented to the customer— highlighting and products used by others in the benefits and opportunities for growth. network, can be analyzed using Value-based approaches are often more sophisticated models to arrive at probable cross-selling opportunities. One relevant62
  • 62. Increased role of data and analytics in cross-selling Figure 1 Predictive Value Social networks Rules Person non-financial example is Amazons 1. Data Mining can uncover potential product recommendation engine that is customers who can be targets for cross-selling, based on users who make similar and lead to generation of off-line offers. purchases. (Refer Figure – 1 for “Increased 2. CRM Systems for sales, marketing and Role of Analytics in Cross-Selling”.) servicing, can use online analytics toBarring the first approach, where the number make cross-selling offers.crunching is done mostly in a persons brain,every other approach calls for heavy use of 3. Predictive Analytics can be used toanalytics—the analysis of data, as well as the make both online and offline offers bycreation of models, rules engines, and offer predicting most likely choices of thedatabases. customer based on past data. Analytics in cross-selling Figure 2 Other technology used in cross-selling includes event Reporting processing, rules engines and more. Text Business Analytics Intelligence Cross- selling Predictive Data Analytics Mining 63
  • 63. Role of Analytics in purpose of cross-selling. Though they Cross-selling may not be part of a suite of products, point solutions are easy to integrate with existing point-of-sale/ service solutions. The role of analytics in cross-selling is Often, these solutions are an easy way of described in Figure 3. bringing cross-selling to an existing environment with minimal changes to Cross-Selling Solutions existing systems. Most of them rely on specific technologies and some rely on a combination of technologies. Examples 1. Home-grown or Assembled Solutions: include Finacle Customer Analytics, Amongst internal initiatives to use Customer XPs, and TIBCOs Cross- predictive analytics, the most common Selling Solutions. application is often cross-selling. In- 4. Channel-specific Solutions: Some house data warehouses provide the data, solutions are designed around specific and business intelligence tools, predictive channels—a call center, for example. These analytics tools, rules engines and coding solutions can monitor call center volumes, provide cross-selling solutions. and trigger extensive cross-selling with 2. CRM Solutions: CRM solutions from incoming calls if the call volume is low. leading vendors—such as SAP, Oracle, When call volumes are high, opportunities etc.—come with cross-selling modules, for follow-up are generated. Similarly, which can be configured and used along outbound call prioritization can be done, with the sales and marketing modules of based not only on probable success rates, the solution. CRM analytics are used to but also based on higher probability of provide the data and power the cross- cross-selling. selling engine, with the operational CRM providing the delivery. Some core Challenges in Leveraging banking solution suites that offer a CRM Analytics solution also offer cross-selling solutions Analytics certainly present a summative view through their customer analytics module of customer transactional and behavioral (for example, Finacle Analyz). patterns. However, the following challenges 3. Point Solutions: These are specific are slowing down the adoption of analytics by solutions that are made for the primary financial institutions: Role of analytics in cross-selling Figure 3 Role Illustrative Examples of Analytics Used 1. Actual process of cross-selling Predictive Analytics, Portfolio Analysis 2. Analyzing past data to uncover trends Data Mining, Reporting, Business and changes in customer preferences Intelligence 3. Measuring effectiveness of cross-selling Reporting, Web-analytics, Channel Analytics64
  • 64. n Expertise: A combination ofLack of and software. This adds to the cost ofdomain knowledge and data analysis implementing analytics models, whichability, a pre-requisite for effective are already considered on the priceyimplementation of analytics, continues side—especially by small and mediumto be elusive. A banking end-user, banking enterprises. In addition, lengthy,though an expert in his domain, interactive database queries and complexoften faces a challenge to interpret analytics scoring processes can congestand analyze the myriad statistics networks and adversely affect databasethrown up by the analytics platform. performance.A data analyst can compile the statistics · Need for Real-time and Advancedquickly, but is dependent on the business Analytics: End users are no longerusers domain expertise to organize content with analyzing historical dataand analyze the data and communicate and understanding past sales in the form the end-user needs it, to Financial organizations now want real-facilitate an actionable decision. time data streaming and analysis thatThe whole process may involve several facilitates on-the-spot business decisions.iterations, resulting in a significant User demands are fast moving fromlag time between data collection and “what happened” scenarios to “whataction and frustration on both sides. may/ will happen” to be prepared with aPredictive analytics, especially, are ready action plan. Analytics models areconsidered a niche realm, requiring expected to answer what will be theextensive training for effective possible outcomes out of action A vs.implementation. action B. This requires high performancen for Clean Data: Statistical· Need analytics models that are capable of real- models are only as good as the data time data analysis. There is growing fed into them. The majority of statistical interest among banks in advanced models not only demand accurate data analytics—though implementation has with the least possible approximations, yet to pick up. (Refer Figure - 4 for but also require that data be scrubbed “Industry Level Advanced Analytics and neatly formatted in a particular Adoption Trends”.) way to ensure quick and meaningful/ actionable recommendations. However, Emerging Trends in the a significant portion of the customer Analytics Field data, maintained by banks happens to be inconsistent and siloed, making it Over the past couple of years, business difficult to meet the formatting standards intelligence—of which analytics are a of analytics models. part—has been catching the attention of financial services industry decision-n· Operational Difficulties: The process makers, who are realizing the need to of deploying sophisticated analytics transform the increased amount of models usually involves accessing available disparate customer transaction data from and/ or transferring data pattern data into actionable information. among numerous machines and Keeping with the growing interest, the operating platforms—requiring seamless following important trends are observed in interoperability of various applications the analytics field: 65
  • 65. Industry-level advanced analytics adoption trends Figure 4 “What are your firm’s plans to adopt the following business intelligence technologies?” Expanding/ Implementing/ Planning to Planning to Interested Not Don’t upgrading implemented implement in implement in but no interested know implementation the next 12 a year or more plans months Reporting tools 31% 31% 12% 9% 10% 5% 2% Data visualization, dashboards 17% 22% 18% 13% 19% 9% 3% Specialized database engines 18% 15% 9% 8% 21% 22% 7% Business performance solutions 16% 11% 10% 11% 27% 16% 8% Decision support solutions 15% 11% 10% 10% 28% 20% 7% Data quality Management 15% 10% 11% 10% 28% 18% 8% Advanced analytics 9% 11% 10% 10% 29% 22% 9% Complex event processing 8% 5% 6% 6% 28% 34% 13% Text analytics 9% 3% 7% 6% 28% 33% 13% 1% In-process analytics 3% 29% 41% 19% 2% 4% Base: 853 North American and European software decision-makers responsible for packaged applications (percentages may not total 100 because of rounding) Source: "The State Of Business Intelligence Software And Emerging Trends: 2010." Forrester Research. May 10, 2010 n Analytics Applications are Packaged business intelligence vendors are in Demand – Business users, especially expected to find great traction. Many financial institutions, are increasingly small to medium-sized banks are leaning demanding packaged analytic towards SaaS models that allow the user applications that are specifically to use the application through designed for online marketing/ cross- affordable monthly subscriptions selling, fraud detection, online credit without heavy IT or manpower analysis, online trading/ investment investments. Small and medium-sized advisory, and others. To date, many banks will leverage SaaS to architect organizations have attempted in-house analytics applications that meet with customization of analytics applications their specific requirements. to meet such specific ends. Such n Open Source Solutions Gain Traction re-architecture may no longer be – Open source analytics solutions are fast necessary with the emergence of eating into the market share of on- sophisticated event-driven/ complex premise solution providers. Apart event-processing products and predictive from low cost, convenience is also a analytics platforms that can support contributing factor—open source these capabilities. solutions can be deployed alongside on- n as a Service (SaaS) Finds Software premise solutions. Open source is Demand with Smaller Banks – SaaS providing an opportunity for recession-66
  • 66. hit organizations to experiment with a features that will support simulation mix-and-match model and acquire using historical data, which helps components of analytics solutions from experimentation before starting the various providers at a fraction of the actual analysis. price. Just as one might assemble spare nInitiatives will Catch Up with Green parts in the backyard, businesses are Analytics Vendors –Initial green efforts toying with the concept of reaching out in the analytics/ business intelligence to best-of-breed open source vendors for field have come from hardware vendors, various phases of the analytics resulting in reduced energy consumption. process—from charting to data Software vendors are expected to enter the crunching, statistic modeling, predicting, market with offerings that will enable and reporting. The soaring sales of companies to monitor their emissions vendors—such as Pentaho and and sustainability exercises. JasperSoft—bear testimony to the growing popularity of open source in the Conclusion analytics field.nMash-ups Make an Entry – Over the Analytics have a key role to play innext couple of years, many analytics helping the banks to increase revenueapplications are expected to be deployed by discovering and fulfilling genuinethrough coarse-grained application customer needs. The pressure to increasemash-ups, which provide a cost-effective sales is even more urgent now than evermeans to embed analytics into business before and the use of online analytics andprocess—without involving major predictive analytics can make the job ofre-architecture work. cross-selling a non-invasive, seamless partnImproving Analytics Literacy – of every customer interaction. PredictiveVendors are realizing that providing analytics provide the much-needed,applications with rich graphical data-based support to cross-selling, whichrepresentations and complex will convert the task of “selling more” intodashboards is not enough to satisfy an act of “fulfilling a customer need” bybusiness users, unless the users have preemption. By ensuring that the cross-sella means of deciphering the output. That is aimed at optimizing value to theis why we will begin to see vendors customer, banks can gain additionalchurning out flexible and user- business as well as customer loyalty andfriendly models with built-in training stickiness. 67
  • 67. Analytics in Financial Services09 SivaramakrishnanAnalytics as a Solution Rajagopalan Senior Analyst,for Attrition Knowledge Services, Infosys Technologies LimitedSwitching from bank to bank requires surprisingly little impetus for many consumers—a slightly higher savings rate, a free bonus offer, or a non-satisfactory customer service call.Years of investments on customer acquisition have left many banks wide open toattrition—existing customers feel unwanted and competing banks appear attractive.To combat this “grass is greener” syndrome, it is critical that banks shift their focus tocustomer retention. This article attempts to provide a solution to customer attritionthrough the application of analytical techniques. because customer defection has become one of Introduction the most illuminating measures in business.The issue of attrition is an area of serious It is the clearest possible sign that customersconcern for the financial services industry. see a deteriorating stream of value from aThough some banks measure customer company. Attrition is more than a number—itdefections, relatively little effort is invested in can hit a bank severely in terms of revenue andretaining customers. This is unfortunate, income growth.
  • 68. Leveraging Analytics to n Determine the root causes of customer Control Customer Attrition churn n strategies that induce churn Define Attrition can be controlled using various reduction methods. The most effective technique is also probably the most straightforward: n Successful implementation of said understand the customer, and leverage this strategies understanding to provide exceptional The above-mentioned approach will help service and maintain a good relationship. drive customer retention to a large extent. This is possible only when a customers The prime challenge is to transform this behavior is known or examined. approach into a practically applicable Analytics provides this window into analytical solution. customer behavior by leveraging the analysis of historical data. In the absence Predictive Analytics of analytics, banks are often reactive rather than proactive—defining strategies only Predictive analytics utilize statistical models when customers are most likely to attrite, that predict the attritional behavior of a so as to win back their confidence. customer by generating risk scores. This is Unfortunately, these situation-dependent followed by profiling customers based on approaches can only help resolve problems reasons for attrition through cluster analysis. temporarily. In attrition scenarios, analytical techniques play a crucial role in assisting The above-mentioned approach can be banks to manage attrition better. Analytics explained by hypothetically applying the transform the business objective into a data- technique on a representative banks driven problem, which can be solved to arrive portfolio (XYZ): at the final solution/ recommendation. n The transactional behavior and other attributes associated with the customer Defining an Analytics are examined, and used for predicting the Approach possibility of attrition. Analytical techniques—such as customer n A statistical model is built on the data profiling and predictive modeling—hold great collated (the transactional behavior and promise as powerful tools to enhance the demographic attributes of the customer retention and to manage the customer) to predict the probability of problem of attrition. In general, the crux attrition of a customer. of determining a possible solution to a problem is to take a structural approach n Deploying this technique not only towards attaining the solution. The concept provides an alert to Bank XYZ stating that of attrition should be viewed in multiple there are customers who are likely to end business perspectives to decide on the their relationship, but also helps in optimal approach. The objective is not only identifying such customers. to retain customers, but also to improve Logistic regression is used for building the profitability and increase their lifetime predictive model through which predictive value. To achieve better customer retention scores are generated for the customers. a sequential approach is required: These predictive scores quantify the risk70
  • 69. quotient of the customer towards attrition The intention of performing this analysis is(i.e., the possibility of a customer terminating to identify the group of customers who havethe relationship with the bank is represented a higher probability to attrite, so that thein the form of these predictive scores). bank can take corrective measures to retain them. This decisive and significantIn the example considered, the predictive information is inferred from the predictedscores would be obtained through the modelequation which is mentioned below. attrition scores for all the customers. In our hypothetical case, a higher score impliesClosure Flag = –1.934e – 01 + 1.850e – 05 higher possibility of customer attrition.Income – 7.074e – 01 No. of accounts – 6.109e– 01 Customer age on book The model performance is also validated using robust validation techniques andIn this case, three variables turned out to be diagnostics. Figure 1 (on the next page)the most important in predicting customer illustrates the models predictive power.churn—Income, Number of Accounts held,and Customer Age on Book. The description In Figure 1, the Kolmogorov-Smirnov (K-S)of the variables is given in Table 1, below. statistic measures the difference between the percentages of attritors and non-attritors ofThe customers who are more likely to attrite the sample distribution. Lift is a measure ofcan be segmented out based on these the effectiveness of a predictive model,predictive attrition scores. The probability calculated as the ratio between the resultsof attrition of a loyal customer (with income (probabilities) obtained with and withoutas a measure of loyalty) is less than that of using the predictive model. A Lift Curve isa customer who has not been using the also used to track a models performance overbanks products/ services for a long time. time, and to compare a models performanceAlso, the number of years a customer has across different samples. The stability aspectspent transacting with the bank (Age on of the model is measured through theBook) has a bearing on the probability of Population Stability Index (PSI) that portraysattrition. The more number of years spent, the stability of the model over time.the lesser the probability of attrition.The number of accounts held by the customer The predicted attrition scores from thealso conveys a similar inference—the more the above model provide comprehensivenumber of accounts held, the lesser the information on the customers behaviorprobability of attrition. towards the bank. If the scores are way too Critical customer variables Table 1 Variable Description Income Annual Income of the Customer No. of Accounts Total No. of Accounts Held by the Customer Customer Age on Book No. of Months Customer has Relationship with the Bank 71
  • 70. high for a highly valued customer, they Step 2: Cluster Analysis indicate a trigger to the bank that if the The clustering procedure is based on the scenario is left unnoticed, the chances are K Means Method of Clustering. In the quite high that it might lead to attrition. K Means method, the algorithm runs on This in turn will drive revenue loss. These an iterative mode. The premise for the predicted scores, along with other available iteration is the assignment of a data point customer information, are used for to each cluster, based on the minimum performing a cluster analysis to segment customers and apply strategies. Euclidean distance from the K-cluster centroids to the data points. The cluster Customer Profiling using centroids become more refined as the Cluster Analysis data points in each cluster change based on minimum distance calculation. For Cluster analysis is used to segment the entire a good clustering solution, the within customer population based on demographic cluster homogeneity and between and transactional behavior. Figure 2 depicts cluster heterogeneity should be high. the steps involved in performing the analysis. In this hypothetical example, Proc Standard Step 1: Data Compilation along with Proc FastClus were used in SAS Step 1 involves the collation of behavioral, to generate clusters, after which profiles demographic and transactional information were created based on the resultant clusters. as mentioned above. The data describing the Five clusters were generated based on the reason for attrition of a customer in the past reason for attrition, and their profiles are is also collated for cluster formulation. described in Figure 3. The reasons for Attrition model performance measures Figure 1 Attrition Model KS and Gains Curve 100% KS & Lift = (%Attritors - %Population) 100% 90% 80% 80% 70% % of Attritors 60% 60% 50% 40% 40% 30% 20% 20% 10% 0% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% % of Population % Non Attritors % Attritors KS Lift72
  • 71. Steps involved in cluster analysis Figure 2 Step 2: Cluster Step 4: Providing Step1: Data Step 3: Evaluation Analysis to Generate Suggestions/ Compilation and Validation Clusters Strategies for of Clusters Customer RetentionCluster profiles Figure 3 Cluster - 1 Demographics: 40+ years Aged Customers, Age on Books is More Than 10 Months, Unmarried, Average Income is US $150K. Products with Highest Transaction Frequency and Amount: Cash credit, Current Account, Loan and Savings. Risk Profile: Low Risk (Based on Risk Score Generated). High Avg Loan Balance, High Overdraft and Loan Relationship Size. Cluster - 2 Demographics: Customers Average Age is 30 years, Average Age on Books is 5 Months, Unmarried, Average Income is US $40K. Products with Highest Transaction Frequency and Amount: Over Draft. Risk Profile: High Risk (Based on Risk Score Generated). High on Over Draft, Highest on Current Account Balance, Most Fees Generated out of Current Accounts, More Risky (Worst Bureau Score). Cluster - 3 Demographics: Customers Average Age is 30 Years, Average Age on Books is 5 Months, Unmarried, Average Income is US $35K. Products with Highest Transaction Frequency and Amount: Cash Credit, Credit Cards and Savings. Risk Profile: Low Risk (Based on Risk Score Generated). High on Cash Credit, Credit Card Balance, High on savings, Highest Fee Generation from Cash Credit. Good Bureau Score. Cluster - 4 Demographics: 45+ years Aged Customers, Age on Books is More Than 20 Months, Mostly Married, Average Income is US $140K. Products with Highest Transaction Frequency and Amount: Credit Card, Current Account. Risk Profile: Low Risk (Based on Risk Score Generated). High Avg Credit Card Balance, High Over Draft balance, High Current Balance and High Relationship Size Across Credit Cards, Over Draft and Current Account. Cluster - 5 Demographics: 40+ years Aged Customers, Age on Books is More Than 12 Months, Married, Average Income is US $160K. Products with Highest Transaction Frequency and Amount: Loan, Savings and TD. Risk Profile: Low Risk (Based on Risk Score Generated). High Loan, Term Deposit and Savings (Highest). 73
  • 72. attrition were examined from the collated clusters, it was found that customers aged data, and the prime reasons were identified 40 and above are less risky compared to by analyzing the attrited customers. The customers of the average age of 30 years. three key reasons which were specified by The low risk customers can be offered a loyalty the hypothetical customers were: program—in the form of special discount offers or reward points that can be redeemed. n Not being recognized as a valuable The customers who are more risky can be customer handled in a different way based on their n staff Unhelpful product usage and transactional behavior. n customer care service Ineffective For example, they could be offered a different product which serves to improve their Step 3: Validation of Clusters creditworthiness. On examining the clusters, The generated clusters are validated in it can be inferred that the risky customers terms of accuracy and practical application. transact more on overdraft and hence, to The clusters accuracy over time is validated avoid any potential risk of attrition, those by calculation of the PSI, which tells the customers could be offered a savings product stability of the generated clusters over time. so as to build a long-term relationship. The Figure 4 depicts the PSI for the hypothetical above proposed strategies would definitely clusters generated. help the hypothetical bank maintain a better Step 4: Suggestions/ Strategies customer relationship. The results of the cluster analysis yielded cluster profiles portraying customer Conclusion behavioral and demographic characteristics. These characteristics are predominantly Low switching costs, a lack of trust, and used for strategic decision-making for deteriorating customer service have pushed customer retention. On examining the many customers away from their banks. Population stability index of clusters Figure 4 Development Validation Cluster PSI #Obs %Obs #Obs %Obs 1 95 25.4% 37 29.6% 0.006423653 2 66 17.6% 13 10.4% 0.038319789 3 62 16.6% 27 21.6% 0.013291666 4 63 16.8% 16 21.8% 0.011107509 5 88 23.5% 32 25.6% 0.001746358 Total 374 125 0.0142 PSI for the Clustering Model = 0.0142 < 0.1, hence, Model Stable Over Time74
  • 73. In the financial services industry, attrition checking accounts, to promoting onlineis a major problem with very real impacts banking and improving customer service.on the bottom line. To combat attrition, Each has one goal in mind: improveproactive banks are turning to analytics to customer retention—a key component forstudy the attrition behavior of customers driving future revenue growth andand to deploy various attrition-battling profitability.strategies. These range from providing free 75
  • 74. Analytics in Financial Services10 Vinay PrasadCustomer Spend Analysis: Unlocking Principal Architect, Banking and Capitalthe True Value of a Transaction Markets Practice, Infosys Technologies LimitedFinancial institutions have compiled a wealth of customer transaction data over the years.When properly analyzed, such data can unlock a treasure trove of predictive information—when the customer will spend, where such spending will occur, and how much will be spent.This article analyzes spend events, the techniques to identify spend events, and the process ofutilizing spend patterns to predict customer spending behavior. To conduct such an elaborate analysis, a firm Introduction must know:Transactional information stored in a financial 1. What was purchasedinstitution is embedded with information that One crucial piece of information missing informs the basis of a spend analysis. Moving into transactional data is the details of the specificthe second decade of the 21st century, a key goods or services purchased. As a proxy, theimperative for banks is to extract this merchant type can be used to determine theinformation and convert it into actionable kind of goods or services the client hasinsights. Imagine the capability to predict: purchased.n customer will take his/ her next· When the 2. When was it purchased vacation Here the transaction date and time isn customer will eat out, where he/· When the important, as it will help in identifying the she will go, and how much he/ she will spend sequence of events. This will also help in identifying the frequency of spend on an customer will spend at the mall· When the particular type of good or service. and at what stores 3. Who made the purchaseIt is no longer enough to know how much thecustomer is likely to spend. Marketing managers Was the transaction conducted by the mainglobally now want to know the “when”, “where” account holder or by one of the dependents?and “what”—details which make the analysis This can help gather some demographicmuch more useful. information on the actual purchaser.
  • 75. 4. Where was the purchase made one cannot group customers purely based on demographics with adequate confidence. The Geo-code of the merchant will be of help in identifying the location of the Spend analytics in this article will focus on purchase (unless the purchase was an e- analysis using Model 2, working from spend commerce transaction). data available mainly in the form of credit card transactions. 5. How much was spent This information is stated on the What is a Spend Event? transaction. Using this information, one can take a A spend event is a set of transactions a number of approaches to define a predictive customer makes to fulfill a need. Customers model for customer-spend analytics. The often make a similar sequence of transactions choice of the appropriate model will be based if they have the same need. For example, on various conditions, specific to the case. consider a customer who has the need to go Two such models are highlighted below. on a vacation: Model 1: Pre-defined Customer a) The customer may book his/ her Segmentation itinerary well in advance (say x days as Customers are grouped based on certain this can vary). demographic data, with an assumption that b) On the day the vacation begins, the people of similar demographic backgrounds customer spends on a taxi (in this case it are expected to behave in a consistent way. is paid by card). The credit card transaction of a customer in a particular group is analyzed to derive a c) The customer checks in at the airport, pattern. Here, the bank looks for similar maybe using the credit card. transactions done within a predefined period d) The customer rents a vehicle, requiring a by a group of customers. Once a pattern is credit card swipe. recognized, any new customer who falls within the group is expected to behave in the e) The customer checks in at the hotel and same fashion. swipes the card. Model 2: Customer Behavior-based f) The customer dines out more frequently Segmentation during the vacation, swiping his card at various restaurants. Transactions are analyzed to bring out similar behavior that has occurred at least a certain All of these transactions would have number of times across the customer occurred in a cluster on the time axis, and population. Once such behavior is identified, the time span across the transactions in the the subset of customers exhibiting such cluster would be fairly consistent across the client base. behavior is analyzed against the rest of the customer base to bring out the discriminating To start the analysis, the initial question that factors. Any new customer exhibiting these needs to be answered is the time period across discriminating factors is then expected to which the data should be analyzed—it could behave as per the identified behavior pattern. be one statement month, multiple statement This approach is to be used when the months, or any other time window based on customer behavior pattern is very secular and the hypothesis being tested.78
  • 76. The next question to be answered considers clustered, they may be converted into athe granularity of the data. The level of spend event by capturing the differentgranularity will depend on the target types of goods/ services bought (based onaudience of the analysis. For example, while merchant type), place and time of transactionanalyzing the data, a bank may not be (if transactions are not aggregatedinterested in expenses related to restaurants. across merchants in a merchant type),This may lead to aggregation of all restaurant- time span across all transactions, transactionrelated expenses during the day as one amount, and merchant details. For theexpense. Overall, aggregation reduces the purpose of Model 2 (customer-behavior-number of records to be analyzed—removing based segmentation), the spend event definesunwanted details. a building block for the construction of aEach transaction can be classified by the spend pattern across a larger time frame.merchants industry. Hence, the transactions Spend events can be classified as irregularcan be associated with the type of goods/ or regular, based on the recurrenceservices purchased at a broader level. The across time windows being considered incustomer-spend across months (statements), an analysis (goods/ services may also beclassified by the type of goods and services used as a proxy for identifying regular andpurchased, leads to the critical “when”, irregular spend events, though one has to“where”, and “what” information discussed be careful, as eating out at a restaurantin previous sections: for one customer may be a regular spend,1. What is the regular set of products and whereas for another, it may be an irregular services a customer spends on one). The identification of the spend2. Where does the customer usually spend event is based on a number of criteria on these goods/ services that the bank has to determine based on the nature of the analysis. For example, if the3. How much does the customer spend on target is to identify irregular spend events any of these goods/ services from the data in Figure 1 (on the next page)4. Who makes the purchase of specific the following factors can be used: goods/ services—the primary card-holder a) Number of transactions per day or the dependent b) Location of transactions5. When does the customer make such purchases—not just the time of day, but one c) Day of the week type of good/ service spend event in relation to Based on these criteria, the bank is able another type of good/ service spend events to identify that: a) The regular spending location is the Identifying Spend Events NY/ NJ metropolitan area b) The number of transactions on a regularIdentifying spend events involves the work day can range from 2 to 3clustering of transactions, using a single- Hence, an irregular spend event (marked indimensional distance measure (based on time blue) is characterized by:gap) or multi-dimensional distance measure(based on time gap and other attributes like a) Number of transactions increased toamount spent). Once the transactions are 8 on 6/13 79
  • 77. An excerpt from a card statement highlighting an Figure 1 irregular spend event Posted Date Payee Address Amount Day of week 6/11/2009 ORB MADAQT ORBITZ.COM IL -107.07 Thursday ORBITZ.COM IL 6/13/2009 SEARS ROEBUCK 1684 WOODBRIDGE -24.99 Saturday WOODBRIDGE NJ NJ 6/13/2009 MAID OF THE MIST STORE NIAGARA FALL -11.09 Saturday NIAGARA FALLSNY NY 6/13/2009 NIAGARA PK CAVE OF WIN NIAGARA FALL -14.58 Saturday NIAGARA FALLSNY NY 6/13/2009 NIAGARA PK CAVE OF WIN NIAGARA FALL -92 Saturday NIAGARA FALLSNY NY 6/13/2009 PETRO #371 WATERLOO WATERLOO -27.81 Saturday WATERLOO NY NY 6/13/2009 KOHINOOR INDIAN RESTUR 716-284-2414 -35 Saturday 716-284-2414 NY NY 6/13/2009 COMFORT INN OF BINGHAM BINGHAMTON -135.55 Saturday BINGHAMTON NY NY 6/13/2009 HKK SUPER SERVICE FLANDERS -21.13 Saturday FLANDERS NJ NJ 6/14/2009 EXXONMOBIL 97360424 WEST HENRIET -15.91 Sunday WEST HENRIETTNY NY 6/14/2009 HOLIDAY INN GRAND HOTEL GRAND ISLAND -136.98 Sunday GRAND ISLAND NY NY 6/14/2009 PILOT 00001701 BINGHAMTON -13.79 Sunday BINGHAMTON NY NY 6/14/2009 DNC SCOTTSVILLE TRVL W. HENRIETTA -12.01 Sunday F W. HENRIETTA NY NY 6/15/2009 ZAHRAS CAFE AND BAKE JERSEY CITY -7.44 Monday JERSEY CITY NJ NJ 6/15/2009 PATHTVM NEWARK BM BW 212-METROCAR -54 Monday 212-METROCARDNY NY 6/15/2009 BLIMPIES JERSEY CITY -6.51 Monday JERSEY CITY NJ NJ 6/16/2009 RELIANCE COMMUNICATION 888-673-5426 -33.58 Tuesday 888-673-5426 NY NY 6/17/2009 TOYS R US #6318 ISELIN NJ -27.76 Wednesday ISELIN NJ 6/17/2009 WEGMANS #032 WOODBRIDGE -32.83 Wednesday WOODBRIDGE NJ NJ80
  • 78. 6/18/2009 TASTE OF INDIA JERSEY CITY NJ -2.14 Thursday JERSEY CITY NJ 6/18/2009 TASTE OF INDIA JERSEY CITY NJ -6.37 Thursday JERSEY CITY NJ 6/20/2009 TASTE OF INDIA JERSEY CITY NJ -7.44 Thursday JERSEY CITY NJExcerpt from a card statement highlighting a shift in regular Figure 2spending habits against data in figure 1 Posted Date Payee Address Amount 9/8/2009 SUBZI MANDI ISELIN NJ ISELIN NJ -50.52 9/11/2009 NJT LIBERTY ST.DLY TV7 JERSEY CITY NJ JERSEY CITY NJ -6.8 9/11/2009 WEGMANS #032 WOODBRIDGE NJ WOODBRIDGE NJ -33.34 9/11/2009 HESS 30215 WOODBRIDGE NJ WOODBRIDGE NJ -25.86 9/11/2009 NFI*WWW.NETFLIX.COM/CC NETFLIX.COM CA NETFLIX.COM CA -18.18 9/12/2009 TASTE OF INDIA JERSEY CITY NJ JERSEY CITY NJ -7.44 9/12/2009 NJT LIBERTY ST.DLY TV7 JERSEY CITY NJ JERSEY CITY NJ -3 9/12/2009 WEGMANS #032 WOODBRIDGE NJ WOODBRIDGE NJ -18.87 9/12/2009 WEGMANS #032 WOODBRIDGE NJ WOODBRIDGE NJ -51.38 9/12/2009 NEW JERSEY E-ZPASS 888-288-6865 NJ 888-288-6865 NJ -25 9/14/2009 TASTE OF INDIA JERSEY CITY NJ JERSEY CITY NJ -8.69 9/14/2009 NJT LIBERTY ST.DLY TV7 JERSEY CITY NJ JERSEY CITY NJ -3 9/14/2009 USPS 33382504929213949 ISELIN NJ ISELIN NJ -12.95 9/14/2009 BHAVANI CASH & CARRY ISELIN NJ ISELIN NJ -28.01 9/14/2009 WEGMANS #032 WOODBRIDGE NJ WOODBRIDGE NJ -14.98 9/15/2009 TOYS R US #6318 ISELIN NJ ISELIN NJ -80 9/15/2009 TOYS R US #6318 ISELIN NJ ISELIN NJ -21.39 9/16/2009 NJT LIBERTY ST.DLY TV7 JERSEY CITY NJ JERSEY CITY NJ -16.25 9/16/2009 NJT LIBERTY ST.DLY TV7 JERSEY CITY NJ JERSEY CITY NJ -3 81
  • 79. b) Location around Niagara Falls, NY Tapping into the Predictive c) Preceded by a booking on a travel site Powers of a Spend Pattern ( which was done 2 days in A customer may be missing a couple of advance. spend events here and there, but generally, Note: The classification of a spend event as all clients belonging to a spend pattern an irregular/ regular spend event is based on should have the same general sequence of the time span across which the data is events, and the time gap between the events analyzed. The same spend event may be should be more or less the same. classified as “regular” if the time span covers To identify a spend pattern, it is multiple years where every summer there are recommended that the bank define an such regular weekend trips. error limit for the time gap, so that two Similarly, if the objective is to identify a shift sequences can be considered similar. If in regular spend events (marked in gray across the difference between related time gaps Figure 1 and Figure 2 on the previous two pages), across two sequences is within the error limit, the bank would look at the following criteria: then they are considered to be part of the a) Merchant segment – In this case, focus on same spend pattern. transportation for results If each spend event is denoted by a “letter”, a b) Location – Being same spend sequence can be thought of as a “word”. To identify if two such sequences are part of a c) Average spend – Look for a significant pattern, the bank would have to use a change sequence alignment algorithm—such as the d) Number of transactions – Look for a Needleman/ Wunsch technique. Here, the significant change user will have to define the weights to be associated with the match and mismatch of Translating Spend Events residues and also with gaps in the sequence. into Spend Patterns This will finally lead to a score for the alignment between the two sequences. A spend pattern is defined as a sequence of spend events observed across multiple The user can also define a limit on the score customers, thus outlining the following: between two sequences, for the two to be a part of one pattern. A set of life-cycle events is denoted a) Sequence of occurrence of such events in Figure 3 and Figure 4 (on the next page). b) Time period between two events For example, Figure 4 shows the sequence of Spend events across the transaction history of a customer are taken to form a spend events across multiple customers over a spend sequence which is associated with period of time. the age or other demographic data obtained There are two patterns — FIC and HG — in from the customers records. Spend sequences the data highlighted in Figure 4. FIC, as a across customers go through a discriminate analysis to identify factors that identify pattern, indicates increase in disposable customer segments with similar spend income and hence, Customers 1 and 3 may be patterns. The customer segment will need more attractive to financial services and to be updated on a regular basis to get a better lifestyle firms. HG, as a pattern, indicates picture of the customer spending pattern. readiness for healthcare products.82
  • 80. Notions to be used in a spend pattern for individual spend events Figure 3 Spend Event Denoted By International Vacation I Domestic Vacation D Drop in Payments to Financial Institutions F Increased Transaction on a Dependent Card C Medical Expenses(Hospital Payments) H Expenses Related to Gym G Example of spend patterns Figure 4 Customer Spend Pattern Customer 1 HFICG Customer 2 HG Customer 3 FICD would be treated similarly, hence Privacy shielding individual spending habits. c) Care needs to be taken in disposing of theNeedless to say, the above analysis can intermediate data created during analysis,be seen as an invasion on customer as it contains customer specific patternsprivacy, and to avoid any breach to (though if data is devoid of customercustomer privacy, the bank should take identity information, linking the twocare of the following: becomes extremely hard).a) Create an inability to link the transactional and demographic data Conclusion back to customer identity information.b) Intrinsically, a repetition of a given Transactional data stored within financial sequence of events is required to form a services firms provides a wealth of pattern, based on which the discriminate information that can be used to better analysis would provide the demographic integrate the customer into the financial information—leading to categorization services firm and the business ecosystem. The of customers. Customers in a category information extracted can provide goods/ 83
  • 81. service providers powerful insights into data, never allowing it to be linked to the customer behavior—driving improved analysis process. The information gathered targeted marketing efforts. from this analysis process should be linked to Care should be taken while doing such demographic segmentation for further analysis to safeguard the customer identity marketing actions.84
  • 82. Analytics in Financial Services11A Dynamic 360o Dashboard: Vishal Gupta Technical Architect, Dr Radha Krishna Pisipati Principal Research SET Labs, Scientist,A Solution for Comprehensive Infosys Technologies SETLabs, Limited Infosys TechnologiesCustomer Understanding LimitedIn todays financial services marketplace, competitiveness hinges on achieving a dynamicview of your customer. A static view of the customer, as a conformed dimension, will nolonger suffice for decision-making on portfolio marketing. This paper presents a high-level,pragmatic approach for providing a dynamic customer 360o dashboard around variousproducts and services in a financial services environment. Analytical engines supporting thedashboard aid in improving decision-making to market the right service. customer—necessary in a market defined by Introduction increasingly sophisticated customers, global competition and innovation.With the advent of Customer RelationshipManagement (CRM) and Master Data This comprehensive view cannot be based onManagement (MDM) solutions and products, scattered solutions. Organization-widethe focus on a comprehensive view of business integration is necessary to achieve an accuratecritical customers has moved to the forefront and actionable profile of the the financial services industry. To sell the Integrating all identified data sources andright products to the right customers, banks systems, and generating derived metrics,rely on information that is not only insightful, requires a sponsorship commitment andbut also dynamic enough to maintain governance approval. The metrics collectedoperational efficiency. must be vigorously checked for quality, completeness, and dependability before theySocial media is a recent emergence that has can be depended upon to provide aprovided additional insights into customer comprehensive customer view.behavior. Online and mobile bankingproducts and solutions have also been useful The presentation of these metrics and the abilityin providing metrics that define customer to dynamically change the parameters and studybehavior using analytics. Such metrics are critical the impact in real-time, will provide a dynamicin building a comprehensive profile of the 360o view of the customer.
  • 83. The Decision Makers achieving a comprehensive understanding Perspective of the attributes that define a potential (or existing) buyer of a product/ To be effective in todays increasingly service. These attributes fall into four o competitive financial services environment, dimensions which sum to create a 360 view a decision maker must: of the customer: Customer Group Dynamics, Business Value to Customer, n Understand frequent changes of Customer Value to Business, and Customer customer interests and taste Behavior/Experience (see Figure 1). n Measure the attraction towards A. Customer Group Dynamics: competitors (attrition) n Understand customer sentiments This dimension is a collection of key metrics related to social, professional and demographic n a realistic view of customer Achieve details of a customer. The categorization is loyalty to the organization dynamic, and changes with every input received na comprehensive view-point of the Attain from internal sources (such as marketing and customer profile updates) and external sources (such as n information (for example, Provide social networks and other media). promotional sales) at the right time, B. Business Value to Customer: using the right touch points Each customer has a perception of the business n Be provided seamlessly integrated value he or she has been provided by various streaming data from ATMs, web events, products or services of an organization. The and other external sources business perspective of this value is provided In short, competitiveness hinges on by this view. Salient points for understanding a customer comprehensively Figure 1 The demographics ? ? created The value and key metrics of for the customer by the customers the organization current group Customer Business Group Value to Dynamics Customer Customer Customer Value to Behavior / The potential and ? Business Experience The customer’s ? actual value till behavior, ideologies date provided by and experience the customer86
  • 84. C. Customer Value to Business: preferences are most valued for potential portfolio growth.Along with the loyalty score of a customer,the value the customer has brought till date, oand is likely to bring, can be presented for a Achieving a 360 View of the Customer: Solutionbetter understanding of the value the Approachcustomer provides to the business. Customerportfolios linked to such information can Defining what one needs to know aboutassist in providing a new dimension of the a customer is one thing, achieving suchpotential impact individual customers may understanding is quite another. The plethora ofhave on the group they belong to. applications residing within an organization,D. Customer Behavior/ Experience: along with the addition of external and socialAll experiences formally captured and information requirements for a comprehensiverepresented in this view can provide view, has demanded that a more customer-information about the customers perceived centric Enterprise Architecture (EA) be built.understanding, and enhance the value-to- What follows is a brief explanation of each ofsatisfaction score dynamically. Customer the layers present in the proposed solutionpsychology, responses to queries and channel (see Figure 2). Logical architecture for comprehensive view of customer Figure 2 Dynamic Layer Rules Engine; Complex Data Processing Interface Alerts and Operational Dashboards Prints and Layer Messaging Reports Downloads EVENT CAPTURE & PROCESSING Data Mining Customer Analytics Analytical Layer Cube Time Series Analysis; Statistical Models; OLAP Integration Layer EDW Subject Marts Master Data EAI; ETL; CDC Data Source Layer Core HR and External Regulatory Feedback and Streaming Data Banking Others Compliance Unstructured 87
  • 85. Data Source Layer forecasting. They are also helpful in determining ongoing business strategies and Transaction systems—such as the core banking solution, cash management, ATMs, Internet measuring the results of decisions. Banking, and credit cards—form the basis for Solutions in the analytical layer will provide the Data Source Layer. Other sources include: various dimensions of customer analysis including n data sources relevant to the bank External the following (Cunningham, et al., 2004): n Unstructured data (such as customer n Customer Voice of feedback and social media information), ? to Ability understand customer which constitutes a major percentage of sentiments data within an organization n Recommendation n respect to regulatory compliance Data with ?to track fast-moving items Ability and provide cross-sell and up-sell Integration Layer recommendations to customers— For the benefit of stream managers and process based on segments according to owners, operational stores and subject marts are demographics, psychographics and essential for tracking performance and relevant spend behavior business measures. The integration of various n Personalization kinds of data sources improves the relevance ?to personalize according to Ability of the information and assists in providing a customer interests and taste o 360 view of the customer. Click stream analysis ? Such integration can be achieved using a nManagement Churn variety of Enterprise Application Integration ? identify the root causes for Ability to (EAI) solutions. For example, Extraction, customer attrition and derive Transformation and Integration (ETL) tools appropriate response strategies are used to integrate data from the Data Source Layer and consolidate it into subject n Profitability Customer marts and data warehouses. Change Data ? determine the profitability Ability to Capture (CDC) is a low latency, non-intrusive of each customer technique to acquire changes (for example, n Analysis Channel insert, update and delete) performed on the ? to evaluate the customers Ability data stored within an enterprise for near real- preferred channels (for example, time tracking of customer transactions. For Web and Mobile), according to banks, this log of sequential event changes on profitability the data affected by transactions, can trigger information events and enable actionable n Segmentation Customer decision-making (for example, in the case of ? to segment customers for Ability fraud). CDC is efficient when operational different analysis activities are running twenty-four by seven, n Loyalty Customer and the business is spread across the globe. ? understand loyalty patterns Ability to Analytical Layer among different relationship groups This layer provides more meaning to the data. n Service Analysis Customer The tools in this layer perform complex ? track and analyze customer Ability to analysis to enable better comparison and satisfaction, the average cost of88
  • 86. interacting with the customer and the into a single, enterprise-wide solution. The time it takes to resolve complaints, etc. portal facilitates accessing, publishing andn AnalysisCampaign downloading the necessary data, and also provides services, such as workflow systems. ? evaluate different campaigns Ability to The advantage of the presentation logic lies in and their responses over time its ability to provide the users requests forn Behavior AnalysisCustomer information retrieval, present the data in a ? to Ability determine fraudulent pre-defined format and provision for ad-hoc transactions queries. Accessing the information is usually ? to Ability determine unwanted performed as a function of business logic. transactional behavior The Dynamic LayernRisk Assessment An add-on to the interface layer, the dynamic ? assess risk associated with Ability to layer provides a single comprehensive view of customers the organizations assets and performance. The customer being the core, the userThe analytical layer will also house the data interface will enable decision-makers tomining component of the solution. Data generate insight from data dynamicallymining is emerging as a key technology for captured and processed when an event isenterprises that wish to expand their encountered. The user/ decision-maker cancompetitive advantage and improve the also be provided with an ability to simulate anquality of their decision-making by event and view the impact in real-time.exploiting operational and other availabledata. This field does not constitute an The changed scenarios from various sourcesindependent research area, but is an will be provided via a common database.interdisciplinary field to address the issue of The data, based on pre-defined rules andinformation extraction from large databases. algorithms, will be processed and the resultsData mining can achieve high return on visualized in an interactive way that isinvestment decisions by exploiting a banks designed or customized by the user or usermost valuable and often overlooked assets. groups. The interface for managing the rules in the engine will also be made availableThe Analytics landscape has progressed from and will assist in simulating a products andbeing merely descriptive to predictive—where service portfolio for a customer or a group ofpossible future states could be predicted based customers, thus enabling customized serviceon current and past events—to prescriptive. In or product offerings.a prescriptive environment, actionableinsights are prescribed to the business for Various design options can be provided toachieving a future state. The perspective user groups at various levels for a comprehensiveanalysis adds value to the dynamic dashboard view of a customer. One such design is illustratedby facilitating proactive decision making. in Figure 3 on the next page.Interface Layer Other dimensions can be added, and a comprehensive view can be provided as aThe interface layer will become the web- combination of the various metricsinterface (or portal) for information logistics, mentioned earlier. The deviated path can beenabling a consistent view for all in the further drilled down to the variousorganizations hierarchy. The portal is based dashboards that can provide a more detailedon corporate business requirements that view of customer experience and can beallow various applications to be integrated subject matter specific, such as: 89
  • 87. A 2-D design of the customer experience Figure 3 Business Value Customer Value 5 E 4 D 3 C Preferred 2 B Actual Group Path 1 A Starting Point n Risk Assessment databases, coupled with myriad social and user generated information combine to n Management Account create a tidal wave of data. Managing this n Performance Management Portfolio data and transforming it into actionable n Service View Integrated information requires the development of a o n The Exception Management modern, 360 dashboard. Properly equipped, decision makers will be able to proactively n Alert Analysis and effectively market products and services n Monitoring Activity to their customers—a competitive advantage for any organization. n and Up-Sell Monitoring Cross-Sell n Money Laundering Fraud and References o In summary, the dynamic customer 360 ° dashboard plugs a financial services decision 1. Herschel, Gareth. “The Customer Data maker into a stream of information—the Every Organization Should Have”, customers experience, the customers value to Gartner Publication. Gartner Research. the business, potential offerings that can be Publication Date: 19 August 2009. made to the customer—which aid in achieving 2. Loshin, David. “Customer Centricity, actionable decision-making. Master Data and the “360o View”. A Dataflux white paper. Conclusion 3. Pudi, Vikaram and Radha Krishna P. “Data Mining”. Oxford Publishers, 2009. Achieving a comprehensive understanding 4. Michiels, I. “Customer Analytics - of the customer is a challenge for any Segmentation Beyond Demographics”. organization. Siloed applications and Aberdeen Group Document, August 2008.90
  • 88. 5. C. Cunningham, Il-Yeol Song and P.P. Chen, “Data Warehouse Design to Support Customer Relationship Management Analyses”. In Proceedings of Dolap 2004, November 12–13, 2004.6. Duxbury, Proteus. “Embracing the Language of the Decision Maker”. Architecture and Governance Magazine, Issue 6-3. content/embracing-language-decision- maker 91
  • 89. Analytics in Financial Services12 Manoj PandeyDeveloping a Smarter Solution Senior Lead Analyst, Infosys Technologiesfor Card Fraud Protection LimitedTo protect customers and organizations from fraudsters, various financial institutionsacross the globe have implemented card fraud solutions. These solutions are designedwith the purpose of preventing losses by using better risk-prediction techniques, keepingtabs on spend management, and managing the customer experience. Maintaining the rightbalance between these objectives is driven by a firms operational risk managementphilosophy. The goal is to retain profitable customers by providing them with a consistentpositive experience. This article discusses various attributes of an effective card fraudsolution and its practical implications. The article also summarizes the concepts relatedto card-holder profiling, advanced analytics, metrics to be tracked, and components of afuture card fraud solution. There are emerging alternatives available Introduction that effectively address fraud. However, due to cost constraints and the current economicLoss related to card fraud for the US payment conditions, banks are reluctant to invest inindustry is estimated by many to be nearly $10 expensive alternatives. For example, “Chipbillion. Though this volume is only a fraction of and PIN” cards cost about US $1.50 each,the total amount of card transactions, it compared to 20 cents for a magnetic stripesignificant considering the unpredictable nature card. Also, the implementation of newerof fraud being committed. For customers, the alternatives often requires high investment inprimary concern is post-fraud (i.e., how fast they infrastructure. Moving forward, the mostare going to get their money back). For banks, the effective and cost-friendly answer to theprimary goal is evolving and implementing fraud problem is developing more advanced,fraud strategies to ensure that fraud does not analytics-based solutions.occur in the first place.
  • 90. Developing an Effective minimal human intervention. Detecting Fraud Solution fraud after it has occurred is not adequate. The ability to flag suspicious activity at the Unfortunately for banks, card fraud is on right time is the key. To consistently achieve the rise (see FinCENs Suspicious Activity this, any solution design goal should be Report [SAR] for Credit/ Debit, Figure 1). three-fold: This indicates that those in the business n Analytics-oriented: Packaging analytics of committing fraud are still able to challenge with a solution can boost performance. existing fraud systems. Leading vendors in Consider the following: the fraud solution space (e.g., FICO, Fidelity, Fortent, SAS, RSA, and others) are ? modules can be integrated Advanced challenged to develop better products to with the product to provide tackle the growing fraud threat. This transaction alerts based on requires investment in product innovation to scenarios—making the solution more stay ahead of fraudsters. It also calls for relevant to the business. solutions that are integrable with enterprise- Transaction ? profiling can be wide modules related to customers, accounts, accomplished based on transaction products, transactions and fraud types. types (ATM, POS, IP Country and Perhaps most importantly, vintage rule and Amount). The profiling process can score-based models need to be updated by include a calibration module to keep integrating various optimization modules. it updated with customer behavior. Developing an effective fraud solution first A sub-module can be added into the ? requires a better understanding of fraud calibration module to help identify dynamics. Fraud is a multi-dimensional normal behavior compared to risky threat that continuously evolves and shifts behavior. This can help connect patterns in response to fraud prevention fraud decisions to customer efforts. Fraud solutions also need to be multi- decisions throughout the service dimensional, advanced and versatile, with lifecycle of the portfolio. SAR for debit/ credit card fraud by depository institution Figure 1 50000 40000 30000 20000 10000 0 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 SAR by Depository Institution94
  • 91. ’Behavioral ? attributes can be n Auto-calibrating profiles used to included in profiles used for analyze online variables and re-scaling. decision-making. This is an n Advanced analytics used to generate enhancement over the activity scores based on vintage profiles of fraud/ attributes used earlier. non-fraud cases.nEnterprise-oriented: The solution The idea is to develop various levels ofshould be able to cover multiple protection in the solution so that fraudsproducts, channels, clients and locations. from different categories can be detected.With multi-dimension capabilities, the The protection layers can be designedproduct will be able to suit the as per the fraud complexity shown inenterprises needs better than an isolated Figure 3 on the next page. Achieving end-to-solution trying to target a specific area or end fraud protection requires significantscenario. improvement in two areas: card-holdernTechnology-oriented: Some of the profiling and advanced analytics, as discussedtechnology pieces the solution can focus in the following sections.on are component-based design, service-oriented architecture, open systems to Card-holder Profilingconnect to multiple applications withoutglitches, high reliability and efficiency. Card-holder profiling forms the foundationThe typical transaction flow in a card fraud on which various other components of thesolution is outlined in Figure 2. solution are deployed. The profiling processOf the modules highlighted in Figure 2, the should be able to discriminate between ascoring engine is perhaps the most critical to normal usage and a risky one. An example ofthe success of a card fraud solution. The the transition from normal usage to riskyscoring engine usually consists of a series of usage is highlighted in Figure 4 on the next page.sub-modules: To ensure a rock-solid base from which tonCard-holder profiling used to track build the rest of the solution on, theprofile variables related to fraud. profiling process needs to evolve from purenAbnormal profiling used to generate segmentation-based static variables toscores based on outlier variables related behavioral variables reflecting the natureto fraud. of risk. Customer segment-based profiling Transaction flow in a card fraud solution Figure 2 Data Model Scoring Business Rule Case Reports Profiles Engine Engine Manager Tuning Engine 95
  • 92. Protection layer and fraud complexity Figure 3 Fraud Complexity Protection Layer Low Identified Fraud Types (from historical fraud data) Transformed Fraud Patterns (old fraud types trying to disguise the system) New Fraud Patterns (new types of fraud activity) Feeble Patterns (sleeper modules getting activated on triggers) High Network Attacks (large-scale attacks on financial institutions) (based on variables, such as customer age segmentation variable; however, when or transaction amount) usually does not investigated further, there can be cases when convey much about the transaction risk low value and high volume transactions are involved. To improve the efficacy of profiling present in a high value transaction segment. efforts, profiles must be created at the The key is to start analyzing the data at a customer level, recognizing that each granular level and then aggregate at higher customers behavior is unique. Also, the levels to find structures and patterns. profiles need to be updated on a real-time The profiling process should also cover basis, based on customer activity. The goal the global nature of fraud. It is usually is to provide the system with the ability to thought that one transaction impacts only compare a customers recent behavior with one profile. Considering the multi- his or her past or risky behavior. dimensional nature of fraud, it is imperative Another profiling problem is the issue of that multiple profiles be read and updated aggregation. Aggregation tends to compare from a single transaction being triggered different transactions on similar metrics. from an ATM or merchant across the For example, purchase value can be a globe. An example could be an ATM-level, Normal usage transition to risk usage Figure 4 Normal Usage Risky Usage Occasional Card Usage High Velocity Electronics and Entertainment Cash Advances Occasional Traveling Cross Border Few Internet Transactions Regular Online Purchases96
  • 93. country-specific profiling process. However,this may have an impact on the customer Advanced Analyticsexperience. As such, its implementationshould be carried out with utmost care and For the new generation of fraud solutions,only when there is sufficient evidence to adapting to a rapidly changing environmentbelieve that the transaction behavior is remains a key challenge—models must beabnormal. adapted and updated to ensure thatThe concept of real-time profiling further historical performance and model weightsenhances the ability of the score to make a are relevant and effective. The previouscorrect decision. With a real-time profiling generation of models had their modelenhancer, it is expected that various model weights frozen and there was no way to knowmetrics (see Figure 5) will benefit. if the model was still performing optimally. Fraud model metrics Figure 5 Metric Description Model Effectiveness A measure of customer assets protected against the risk computed by the formula: (Total Customer Asset Protected – Gross Loss)/ Total Customer Asset Protected. Rule and False FPR is a model metric which tracks number of cases to be Positive Ratio (FPR) investigated to confirm one fraud. A typical FPR maintained in the industry is <=20. Various rules are designed in the model to capture frauds. A rule-based FPR metric will highlight the weak rules not able to capture frauds. Customer Activity Multiple rules can be designed to tackle frauds occurring for one and FPR type of customer activity. This metric can track customer activities that are vulnerable to frauds. Bench-marking Bench-marking against internal/ external systems on common metrics can further help in understanding the risk management and mitigation. Cost of Fraud This metric tracks the cost of fraud detection, which can Detection complement the FPR metric. Fraud Cases This metric reports the number of fraud cases captured against a Captured target set by the business. Model FPR This metric captures the model level FPR against a target set by the business. Model FPR Vs. This derived metrics shows model efficiency against its Model Effectiveness effectiveness. Very useful for managing customer experience and fraud. Top 10 Riskiest This metric risk ranks the fraud activities that can help optimize Fraud Activity the rule designing process. (Based on Dollar Loss) 97
  • 94. A common metric, Points to Double the such a capability are quite significant). Odds (PDO), can be used to measure score An analytics-driven, Advanced Scoring performance over time. A rescaled score Technique can supplement the Shared distribution tends to be more stable and Network Repository knowledge with real- realistic compared to an originally developed time updates from the case management one. The rescaling is done to focus on the system. The base score can be augmented changes in behavioral patterns, rather than with internal and external sources to the fraud pattern. For example, one would provide a complete and updated view of expect to see a changed pattern during the the risk. The final score from the above following scenarios: process is fed into the business rules designed n of Apples iPad - Customers Launch specifically to capture current fraud patterns line up to buy electronic gadgets. (which are not captured by the Advanced Scoring Engine). n Christmas Celebration - Customers spend on buying gift items during To monitor the performance of the Christmas and New Year. Advanced Scoring Technique, it is suggested that detection rates are plotted n in Interest Rates - Customers Changes against FPR with and without the changing their spending patterns technique (Figure 7 on next page). Detection accordingly. rate is the percentage of frauds correctly n in Inflation Rates - Customers Increase identified by the model against the total changing their spending patterns actual frauds. As expected with advanced accordingly. scoring, the detection rate is almost 6% The model (see Figure 6) should have higher at an FPR of 10. This is significant the ability to measure its performance in considering the value at risk. The goal of the real-time and adjust the model weights Advanced Scoring Technique is to maintain accordingly (though a technological and improve model performance on a challenge, the benefits to the business of continuous basis or between model refreshes. Advanced scoring engine Figure 6 Base Dynamic Advanced Scoring Transaction Base Score Engine Model Threshold Review Shared Network Update Model Threshold Repository Weights False Positive and Review Case False Negative Table98
  • 95. Plotting advanced scoring performance Figure 7 70 Advanced Scoring 60 Detection Rate (%) 50 40 30 20 10 0 0 10 20 30 40 50 FPR (%) WITHOUT Advanced Scoring Technique WITH Advanced Scoring Technique and ATM provides a holistic view of the Other Focus Areas transaction. Features like dynamic profiling can further enhance system efficiency andThe ability to multi-profile is another key reduce latency. This enables the financialfeature of a robust solution. Multi-profiling institution to remain current on the riskfor customer, account, location, point-of-sale scenario and focus on long-term goals, rather Model implementation strategies Figure 8 Fraud Detection LOW HIGH (A) Real Time transaction scoring and decisioning Model Effectiveness LOW Decision Latency HIGH (B) Score based on historical behavior. No real time decisioning HIGH LOW (C)SScore sent to analyst in batches. Delay in transaction and decisioning LOW MEDIUM HIGH Model Complexity 99
  • 96. than responding to risk on a daily basis by n Rules: A set of rules packaged Implicit adopting course correction. with the solution and driven by the Shared Network Repository. These can be The solutions implementation option is updated on a preset frequency (hourly/ another area which should be evaluated in daily) by learning in real-time from fraud detail. The choices span from real-time, occurring at other institutions and in online solutions to batch processing other geographies. solutions—each option varying in model effectiveness, fraud detection, and decision n A case investigation drill-down capability based on customer, account, transaction latency (Figure 8). and other user-defined attributes. Preparing for the Future n Dashboard Alerts: Alerts based on specific scenarios such as non-monetary account activity, card block and In the future, it is expected that new fraud replacement, high login failures, and patterns will emerge—each calling for more external alerts. complex and effective solutions. Some areas that need to be addressed by future fraud n Profiles extended to merchants, Model detection solutions include: devices, accounts, customers, and transactions. n Alert System: Ability to queue Advance alerts based on rules, user-defined n Customer-based scorecard (customer life- priority, transactions and other relevant time value, cross-sell/ up-sell, and parameters. profitability) inclusion in the solution. Managing customer and fraud Figure 9 Balancing Customer Experience and Fraud 60 Detection Rate (%) 40 With a better model in place, the 20 bank can release over 0.5% more with a detection rate of 40%. 0 0 0.5 1 1.5 2 2.5 3 False Positive Review Rate (%) Detection Rate (%) - V1 Detection Rate (%) - V2100
  • 97. n of Foreign Assets ControlOffices Solution vendors also need to clearly(OFAC) watch list integration in the rules position their products—targeting key pain-engine. points of the bank.n level cases and investigation.Customer ReferencesnCase linking capability.nSAR evidence analysis. 1. “The SAR Activity Review – By thenmetric dashboard.Model Numbers”, Issue 13, Exhibit 5, Jan 2010, FinCEN.Apart from stopping frauds, anotherimportant aspect of a fraud solution is 2. “Reimbursement Tops Customermanaging the customer experience. This is Concerns Following Card Fraud”, Tellerusually achieved when the solution is Vision, Dec 2009, Issue 1388, to treat each customer uniquely and 3. “Emerging alternatives to chip and PINis able to balance fraud and customer to tackle card fraud in the US”,expectations as per defined business MarketWatch: Global Round-up; Octpolicies (refer to Figure 9). Sometimes 2009, Vol. 8, Issue 10, p197-198.strategies which are purely fraud-orientedbackfire—creating more problems for the 4. Crossley, Jane, “Credit card fraud: Howcustomers than expected benefits. Fraud to fight it”, Credit Management, Maysolutions are designed with the purpose 2009, p26-27.of preventing losses by using better risk 5. “ABA Offers Debit Card Fraudprediction techniques and also keeping a Prevention Tips”, Teller Vision, Martab on spend management. The goal is to 2008, Issue 1367, p6.retain profitable customers by providingthem with a consistent positive experience 6. Graves, Brad, “Falcon Helps Spotwith low FPR. Credit Card Fraud”, San Diego Business Journal, Sep 2004, Vol. 25, Issue 36, p9. Conclusion 7. Flynn, Elizabeth, “Working together to combat financial crime”, MoneyAs the digital economy evolves, new Management, Jan 2008, Vol. 22 Issue 1,methods for committing card fraud are p11.on the rise. Though compliance remains 8. Knapik, Jaroslaw “Using Technology tothe top priority for banks, there is a need Combat Financial Crime in Retailto devise strategies around risk categories. Banking (Strategic Focus)”,Particularly for card fraud, banks should Datamonitor, Dec 2008.focus on advanced concepts such as dynamicprofiling, advance analytics, fraud metrics,and implementing technology projects toincrease the transaction processing speedand accuracy of fraud detection. There is aneed to balance the cost of the fraud solutionand the benefits it is driving for the business.With increasing costs, banks can adoptstandardization in their business processes. 101
  • 98. Analytics in Financial Services13Using Adaptive Analytics to Jehangir Athwal Lead Scientist, Larry Peranich Principal Scientist, Scott Zoldi Senior Director, Analytic Science, Analytic Science, Analytic Science,Combat New Fraud Schemes FICO FICO FICOThe fight against payment card fraud resembles an arms race, with card issuers deploying evermore sophisticated anti-fraud measures and fraudsters continually evolving strategies toevade those measures. Issuers typically rely upon neural network fraud models that takeadvantage of huge historical datasets to recognize recurring fraud patterns and reduce fraudlosses. However, the fraudsters decentralized nature and short time-frame give them anevolutionary advantage over the issuers multi-month to multi-year analytic development cycles.Adaptive analytics, when used with these neural network models, swing the advantage back tothe issuers by continually adapting the fraud detection models based on the latest fraudbehavior. This not only improves model performance, but also extends the useful lifetime ofthe static neural network models. In a test described in this article, adaptive modelingtechniques improved fraud account detection by nearly 20% and real-time value detection bymore than 15%, at a 10:1 Account False Positive Ratio. One way to boost the performance of fraud Introduction models, and sustain the performance longer, is to add a second layer of analytics that isSuperior fraud detection is traditionally based adaptive or self-learning. Adaptive modelson utilizing an abundance of historical bring the perspective of the present and neartransactional and fraud reporting data to build present—what is currently happening in theand finely tune neural network models and portfolios operational environment—to fraudother advanced analytics. But financialinstitutions are always looking for ways to detection.combat newer fraud schemes—schemes that This article discusses the use of adaptive modelsarise between fraud model developments and in a “cascade” architecture, where they act asare not well represented in the historical data. a secondary analytic layer. Benefits includeAdditionally, changes in regulatory constraints, faster interception of new threats, more robustpayment card technologies, and business custom models, better fitting consortiumoperations can modify the characteristics of the models, flexibility to increase focus ontransaction stream. This modified transaction particular types of fraud, and reduced riskdata can look suspicious to fraud models based when deploying new detection approacheson historical data, raising false positive rates. into production environments.
  • 99. In a cascade architecture, as shown in Cascading Fraud Scores Figure 1, the adaptive model is invoked after the base model has generated a fraud score. Adaptive models deployed in a cascade For transactions with base model scores architecture strengthen conventional above a specified threshold, the adaptive fraud detection by providing an additional model generates an additional score. This method of refining the prediction of adaptive score is then combined with the future normal and fraudulent behavior. original score into a blended score, which The cascade architecture combines detection refines the prediction of fraud and based on historical customer transactional determines whether or not the transaction behavior with detection based on the current is referred for investigation as possible fraud. behaviors observed in the operational environment. At the same time, it keeps the A tight feedback loop from the case technologies separate, which is advantageous management systems used by analysts back both for detection and operations. to the adaptive layer is essential. Data should be captured from case dispositions as they FICOs work with client data shows that occur, populating the fraud/ no-fraud tables the cascade approach delivers significant used by the adaptive model with the model fraud detection lift when used with a base input variables derived from those selected layer model consisting of a strong neural transactions and their corresponding card network built from an abundance of profile variables. reliable, high-quality historical data. A Focused Task Adaptive models benefit large issuers with enough historical data to build a One of the reasons the cascade is able to custom base model tuned to their portfolio increase the proportion of fraudulent characteristics, as well as issuers who transactions in referred transactions combine their data into a consortium (high-scoring transactions), while lowering dataset and produce a single consortium the incidence of false positives, is that the base model that has been exposed to the adaptive models job is very focused. widest variety of fraud types. It analyzes a relatively small set of recent Adaptive models work with supervised Figure 1 models to catch more fraud104
  • 100. transactions and fraud/ no-fraud behaviors Adaptive models also enable companiesto make a determination of the likelihood to concentrate more detection power onof fraud. specific fraud categories. Cascaded adaptive models can access data from multiple fraud/The base model scores all transactions no-fraud tables, thereby directing a tightwith analytics shaped by at least a analytic focus simultaneously onto severalyears worth of historical behavior fraud types. For example, in addition to a(occurring, in the case of consortium general fraud table, there might be amodels, across many banks). It “sees thebig picture”. The adaptive model, in table devoted to card-not-present fraud andcomparison, is focused. It examines only another to cross-border fraud.transactions within a high-scoring range The adaptive model frequently retrainsusing analytics shaped by the present itself in production based on the changingand near-present (e.g., last day, week, etc. data in its fraud/ no-fraud tables. Thisof referred fraud case dispositions, adaptation occurs through the mechanismeither fraud or no-fraud). This enables of dynamic variable weighting. Every timethe model to concentrate a large amount a referred case is disposed by an analyst,of analytic power on a specific task. new data enters the fraud/ no-fraud tables.In fact, the adaptive model’s concentration The adaptive model then adjusts theon high-risk transactions affects how it weightings of its predictive variables basedoperates, particularly its sensitivity to on this new data.changing fraud patterns. The memory size Combination and Separation—Advantagesof the fraud/ no-fraud tables, populated of Model Layeringwith case disposition data, determinesthe number and range of no-fraud and A cascaded adaptive model is an efficientfraud patterns used in adaptive scoring. way to boost fraud detection because itA table storing the last three months works with the base model withoutof cases will capture more fraud examples interfering with it. The adaptive layer canthan a table storing the last week of cases—but simply be “bolted onto” an existing neuralit will also be less sensitive to the most network model. The base model remainsrecent changes in fraudulent behavior. untouched—it is a finely tuned neural network, which best leverages massiveGiven a highly predictive base model, amounts of historical transactional andshorter adaptive memories are preferred, reported fraud data—continuing to produceas they make the adaptive model highly highly predictive scores.sensitive to change, and thus able to detectnew fraud attacks and emerging fraud trends Keeping the layers separate also enablesquicker. Quicker detection minimizes the companies to turn up the “sensitivityimpact of these new fraudulent behaviors dial” of their fraud detection solution toon banks and their customers. However, short-term changes in fraud behavior,shortening the time-frame too much can while shielding the base model from thereduce the data available for training the effects. Adjustments that are valuableadaptive model, reducing the examples of for detection today might not be relevantfraud it has seen and the statistical next week. For example, an adaptivesignificance of its score. A time-frame of model using a table that captures just thetwo to four weeks is usually optimal for a past week of fraud/ no-fraud case data willgeneral adaptive model. be very quick to adjust scoring to fraud 105
  • 101. burst behavior. An adaptive model with a n Flexibility to increase analytic longer time horizon might provide broader scrutiny on certain fraud types. protection across more fraud patterns, Companies can put more focus on rising where an issuer is less concerned about fraud types, without reducing the burst activity. effectiveness with which their base model detects all other fraud, or otherwise Benefits of Adaptive Models skewing it in undesirable ways. n risk when deploying new Reduced When financial services institutions add a detection approaches. Supervised cascaded adaptive model to their current fraud models are highly refined fraud detection solution, they benefit in analytics that undergo rigorous and numerous ways. exacting retraining procedures based on previous years data. The use of an General Benefits adaptive analytics layer provides a low- n fraud detection performance, Higher risk, more forgiving means of trying particularly at low levels of false approaches not fully supported by positives. Performance rises not only historical data. Companies can leverage because a second layer of analytics is being variables in the adaptive model (even applied to high-scoring transactions, but variables suggested by observation or because the adaptive model is scoring hunches), knowing that the variables these transactions based on analyst ultimate weighting will be performed by fraud/ no-fraud feedback on similar the model itself through a completely recent cases. Moreover, by increasing data-driven process in the production the proportion of fraud in referred environment. transactions, the cascade enables analysts Additional Benefits for Companies using to work more fraud cases, subsequently Custom Fraud Models feeding more fraud examples back into the adaptive model’s fraud table. n Decreased rate of degradation. After deployment of a custom model, as fraud n time to detect and intercept Shorter behavior changes, the effectiveness of the new fraud. The adaptive cascade enables model gradually diminishes, particularly the overall fraud solution to adjust to the when the model sees fraud behavior that operational environment as captured in had not existed in the banks historical analyst dispositions of suspicious cases. data prior to model retraining. Fraud detection is thereby able to leverage Degradation will occur much more new variable relationships and fraud slowly, however, when there is an adaptive behaviors that were absent from or not layer absorbing current fraud case data sufficiently represented in the historical and adjusting scores based on new fraud data used to train the base model. New behaviors and patterns. fraud types receive higher scores much sooner than they would without the neffective extension of existing More adaptive layer. Alerted to rising threats, fraud detection to acquired portfolios. companies can also move faster to Financial institutions that acquire implement new fraud rules and other additional portfolios can now consider policy/ program changes necessary to protecting these new accounts with their increase protection and mitigate losses. existing fraud detection solution. While106
  • 102. the characteristics of the acquired and normal behaviors experienced portfolio will differ from those used to across multiple institutions and markets train the supervised fraud model (based over the past year. Companies that use on the original portfolio of historical a cascade architecture benefit from data), the adaptive layer will adjust continuous retraining of their adaptive variables and weights in real-time to the model as it self-adjusts to current fraud/ new production environment and the no-fraud cases. behaviors in the acquired portfolio.Additional Benefits for Companies using Adaptive Models in ActionConsortium Fraud Models As discussed earlier, adaptive analyticsnBroad fraud-type detection with tighter provide value by continually adaptingportfolio fit. Companies that use fraud detection models based on theconsortium models are able to detect a latest fraud behavior. To test this, as wellbroad range of fraud types, thanks to the as to quantify performance improvement,vast amount of data from many different FICO applied the technology to afinancial institutions used to build the UK consortium fraud model at a timemodels. With an adaptive layer, they can when fraud behaviors were in flux. Theenjoy this benefit, while making fraud UK credit card environment underwentdetection more responsive to the specific significant changes in 2004 with thecharacteristics of their own portfolio. introduction of chip-PIN authentication.nRetraining that goes wide and deep. Counterfeiting, which had been the mostWith each retraining, consortium models prevalent fraud type, declined, while cross-are refreshed with data about fraudulent border and card-not-present fraud surged. Performance improvement with adaptive model layer Figure 2 107
  • 103. In this test, FICO took a UK Consortium As seen in these plots and corresponding FICO™ Falcon® model trained on pre-chip- tables, the adaptive technology was successful and-PIN historical data, and ran it on in boosting the overall performance of production data from after the February UK credit card fraud detection in the 2006 mandatory chip-PIN implementation. operational range of the model. The The cascade adaptive model was simply model provided the largest lifts at very “bolted onto” the base Falcon model, low levels of false positives (highest using already computed profile variables scoring transactions), consistent with without any manual tweaking of adaptive the models ability to increase the model parameters or adding of new adaptive concentration of fraud among referred model variables. transactions. At a 10:1 Account False Positive Ratio, for instance, adaptive The fraud detection performance of the modeling improved fraud account base model with the adaptive model vs. detection by nearly 20% and real-time the base model alone is shown in Figure 2. value detection by more than 15%. The horizontal axis measures the Account False Positive Ratio (AFPR), which expresses the number of accounts Conclusion identified incorrectly as fraudulent for each actual fraud account the model An adaptive fraud model, when used identifies. Account Detection Rate in a cascade architecture with a static (ADR) is the number of correctly fraud model, can substantially improve identified fraud accounts expressed as a payment card fraud detection. It further percentage of all fraud accounts. Value differentiates between fraudulent and Detection Rate (VDR) is related to the normal behavior among the most risky amount of money saved as the result of transactions, particularly as fraudsters correct fraud prediction and is expressed change their behaviors to evade the as a percentage of the total amount of anti-fraud measures the card issuers fraud dollars. The VDR statistics presented have in place. As the deployment of here assume that the fraud detection self-adjusting, continually retraining system is in real-time mode, in which case, models grows, financial services companies the model responds to the current can realize additional performance gains transaction authorization request by and value throughout their analytic generating a fraud score before the operations. transaction is completed.108
  • 104. Analytics in Financial Services14To Fight Fraud, Connecting Scott Zoldi Senior Director, Kyle Hinsz Senior Scientist, Analytic Science, Analytic Science,Decisions is a Must FICO FICOFinancial institutions provide customers various types of accounts and an ever-increasingnumber of ways to access these accounts. Traditional fraud detection systems are highlyspecialized to compute risk for a particular access method, but this approach is running upagainst serious limitations. Future systems must take the next step in fraud detection byprofiling a wider range of information and connecting it across customers and accounts.Advanced analytics must not only intelligently use data across these diverse account andaccess combinations, but dynamically detect new fraud patterns with accuracy, even asservice/channel usage and other customer behavior change. a view across the channels or across customers Value of Connected Decisions accounts makes the resulting fraud detectable.Within financial institutions, there is an Another important aspect of connectedincreasing need to connect decisions decisions deals with "know your customer”.across silo decision areas. There is almost no When decisions are made without sharingpart of a financial institution that cannot information across lines-of-business, thebenefit in some way by shared information with results can be embarrassingly contradictory.other areas. For example, originations may decide to offer a new credit line to a customer at the same timeMany fraudsters have determined that as the risk management team is reducingwhile some silos are well protected, the his or her existing credit lines. Or, a credit lineconnections or interplay between silos may optimization strategy may increase thenot be protected. For example, a large inflow credit limit of a credit card, while in aof dollars to a long-existing Demand different part of the business, the fraud team isDeposit Account (DDA) may be viewed as a investigating the card for first-party fraud.normal or a low-risk event. However, if this eventis followed by multiple transfers to newly The value of connecting decisions is notestablished DDA accounts, and subsequently lost on the financial industry. According to aby multiple ATM withdrawals, then this TowerGroup survey of consumer lenders, 80%cross-channel set of transactions points to a of respondents believe that an “integrated viewlarge fraud event that may not be captured of customer data” will become increasinglyby each respective silo solution. Only important in their decision process.
  • 105. Challenges of Connected that need to be addressed. Building a model Decisions segment for every possible combination would rapidly become unwieldy. The first hurdle that must be overcome to In addition to the growing number of realize connected decisions is to implement services, many of the expanding areas in a data model that facilitates the connecting retail and card operations are also doing of information across access channels and business in environments of pervasive, accounts. This framework is extremely unremitting change. Some of these changes important because it is the basis for how are coming from the banks themselves, as both analytic models and human users are they adjust policies to deal with concerns able to manage information from a variety such as delinquencies and regulatory impacts, of sources. Without this foundation, and as they introduce multichannel services institutions find themselves with a variety to attract and retain valuable customers. of ad hoc processes to try to share information, each developed at a different These intersecting changes reduce the accuracy time, by different people, with a different of existing fraud detection methods. For problem in mind. The build-up of these instance, as banks modify their originations one-off solutions results in a system that is policies to mitigate the profit impact of unnecessarily complicated, unreliable, and new regulations, they are reshaping customer time-consuming to maintain. population segments. The characteristic transactional behavior patterns of these Once the data model is set, there are further segments will differ from those found in challenges to introduce the appropriate historical data. This limits the reliability analytic technology. Accurate fraud detection of historical data for building rules and becomes particularly difficult with the complex supervised models. proliferation of bank services and account access channels. Traditionally, fraud analytics Smart, Dynamic Profiling are based on fixed characteristics, like Technology the assumption that all customers will have similar account set-ups and account The first step to enabling connected relationships. Today, however, there is a decisions is to apply the appropriate much wider range of possibilities, and data model to associate customers to models built with such assumptions are less their different accounts. Without easily effective at detecting fraud in populations being able to link account activity back that have many combinations of services and to the customer performing the actions, channels. Moreover, as new services and there is a constant struggle—and huge access channels are rolled out, typically there overhead—to fully understand customer arent sufficient amounts of historical data behavior. in these areas for statically valid analysis or This linking of account activity to a supervised model building and testing. customer can be handled by utilizing the The applicability of fixed characteristic proper transaction header attached to each models can be expanded by building out of the application areas base data formats. additional model segments to cover common For each transaction sent to the scoring combinations of services. However, as the system, the transaction header identifies the number of retail banking services and Customer ID and Account ID associated channels increases, so do the permutations with the transaction. These identifiers are110
  • 106. utilized to fetch and store profiles associated These profiles can be applied to entities suchwith the customer and account. as a card number (PAN), an application number, or an ATM terminal identifier, toThe Customer Profile name a few. As an example, the ATM profileThe customer profile tracks customer may contain variables tracking averageactivity and account summary information withdrawal amounts, aggregations of cardacross all the accounts held at the bank. scores at an ATM device, or a balance inquiryThe customer profile contains pertinent transaction followed by a large dollar amountfeature detectors (variables) associated across many debit cards over a short period.with the customers behavior, and could Figure 1 shows an effective data model andinclude customer location information, this dynamic profiling method that a bankrecent application information, bureau can use to connect decisions. Each recordinformation, and cross-account information utilizes a transaction header that identifiesthat track pertinent ordered behavior between the type of transaction, as well as theone or more accounts. The customer profile associated Customer ID and Account ID foralso includes account summaries, which are profile access. It illustrates the multi-profilingsuccinct summaries of the current state of approach where for a particular transaction,each account held by the customer. Account both a customer profile, which providessummary variables would contain fraud risk summaries of all account profiles for theinformation associated with the account, customer, and an account profile arereason codes for the risk assessment, select retrieved, based on the Customer ID andvariables from the full account profile, and Account ID associated with the currentsummary of recent transaction information. transaction.The Account Profile In addition, channel-specific profiles may be retrieved based on the Transaction RecordThe account profile tracks customer Format, also known as the “body”, whichactivity associated with a particular account. follows the transaction header. For example,The account profile contains pertinent channel silo profiles for a credit cardfeature detectors (variables) associated with transaction could include keys—such as PANthe account, such as account status (number embossed on the card), Paymentinformation, open dates and locations, Instrument ID (unique plastic identifier), andaccount application information, monetary merchant identifier.transactions, as well as cross-channelinformation that tracks activity across The fraud risk score may be based on thedifferent channels (such as a large online channel silo profiles, the account profilebanking transfer into the DDA account (which contain summaries across all accessfollowed by large ATM withdrawal using a channels to the account) and the customerdebit card). The account profile also contains profile (which contain summaries acrosschannel summaries of the activity within all accounts held by the customer).each access channel to the account (for Furthermore, this score can be actioned atexample, a debit card or online banking the transaction, account, or even customerchannel into a DDA). level. This allows banks to customize their fraud actions based on the type of risk—forThe Silo Profile instance, combating merchant fraud atThe silo profile contains highly focused the transaction level and investigating first-features to boost performance in the silo area. party fraud at the customer level. 111
  • 107. Data model and dynamic profiling enable connected decisions Figure 1 Adding Analytics for Multi- behavior patterns of sub-populations of channel Fraud Detection customers while in production. It also continues to learn from ongoing transactions Once a data model is set, connecting and fluidly encompasses new offerings as decisions requires an analytic technology banks bring them to market. that performs well in dynamic, multi- channel banking environments. Self- Learning New Fraud learning outlier analytics meets this need Behaviors “On-the-Fly” by examining ongoing transactions and automatically determining which customer Self-learning outlier analytics, as with behaviors lie outside of the current range of other methods, depends on the crafting of usual activity for account-holders with highly predictive variables, a process that similar characteristics. requires extensive domain and fraud expertise. This technology differs, however, In retail banking, outlier analysis can in the way it learns typical and atypical be performed in a manner that is highly values from these predictive variables. specific to any number of customer segments with different combinations of Traditional supervised fraud models services (such as credit lines, deposit learn these values by examining large accounts, and loans), products (such as quantities of historical data—a process credit cards, checking, and home equity guided by analytic developers. Self-learning lines of credit) and channels (such as teller, outlier analytics infers predictive variable ATM, online banking, and mobile banking). values—on the fly—as it processes production Yet it is also very manageable, since there data. Using anomaly detection and is no need to try to envision all possible other statistical techniques, it dynamically combinations prior to deployment. The scales variables, estimates variable analytics automatically discovers characteristic value distributions, and determines what is112
  • 108. an outlier through comparison within the credit card. It may not be an outlier, or at leastpeer group. much less of one, in a segment where customers have three credit cards with a totalSince this type of analytics does not rely on line of credit greater than US $40,000 and lesshistorical data, it can perform on customer than 20% utilization of the credit line.segments that would have insufficientdata for standard supervised model training, New peer groups can form spontaneouslyor in areas where the historic data has as banks introduce new services, productsbecome obsolete. The analytics learns from and channels. When configured to recognizetransactions in production, adjusting to an external indicator, such as a new segmenthow behavior patterns change as customers ID, the analytics will notice the previouslybecome accustomed to new services over unknown ID, and automatically create atime. For example, typical behavior within new peer group of customers who are usingsegments of customers using mobile this combination of services, products andbanking is likely to change drastically in the few years. Self-learning outlier analyticsaccommodates such changes by constantly The result of this sophisticated technology iscomparing peer group activity and a very straightforward benefit for customers:recalibrating ranges of characteristic behavior A better user experience with new andas necessary. potentially riskier banking methods, and continued self-learning within existingPeer group definition—based on subscribed banking methods. This occurs in two ways:products and services, channel usage, andother characteristics (such as credit risk or 1. The analytics improves and acceleratesaccount balance levels)—is essential to this the identification of fraudsters andprocess. A US $2,600 transaction may be a vulnerabilities threatening new orstrong outlier in a segment where customers existing bank offerings through thehave a single line of credit, say a “student” constant re-evaluation of outlier values. Overview of how the technology works Figure 2 113
  • 109. Early adopter transactions and accounts makes the fraud model highly sensitive to are better protected where sparse or no the most recent fraud feedback. historical data exists. By dynamically adjusting predictive 2. False positives (legitimate transactions variable selection and weighting based on scoring high) are greatly reduced because recent fraud and non-fraud case dispositions, the analytics learns from the behavior of this technique provides an automated way customers within segments and quickly of temporarily turning up the “sensitivity assigns the appropriate risk levels. dial” on specific behavior patterns. The adaptive layer also has the advantage that When used with a traditionally trained neural the outlier model weights will be adjusted network model in an online banking channel, regularly based on changes to the fraud and self-learning outlier analytics lifted detection non-fraud tables that are driven by analyst accuracy by 25%. Where only rules or simple and fraud feedback. models are currently employed, the lift from adding this technology will be even greater. Conclusion A Cross-Channel View of Fraud Complexities mount as banks look to connect information and decisions in a business Self-learning outlier analytics catches fraud environment that is more volatile and less that is invisible to current methods, which predictable than ever before. Changing are often focused on service-, product- or economic conditions, regulations and bank channel-specific “silos”. The complete picture policies, services, and channels are causing is visible to self-learning outlier analytics, customer behavior and fraud schemes to since the characteristic variables compared change as well. If banks do not keep pace, fraud within a peer group can include cross-channel losses will grow at a time when maintaining and cross-account profile variables. profitability is already a challenge. The dynamic profiling technology discussed The maintainable, cost-effective approach earlier compresses transactional data into for customer-level analytics is to lay a solid predictive fraud variables stored in a highly foundation with an effective data model, efficient form for real-time analysis. Profiles and then to generalize existing fraud can capture complex behavior patterns— detection methods by adding self-learning including the sequence, timing and velocity outlier analytics and other dynamic learning of transactions—occurring across different technologies. These techniques also serve as services, products and channels. an excellent means to automatically calibrate not only to changes in the business, The Ultimate in Dynamic Fraud Detection but also to changes in how customers behave, and use channels and services The state-of-the-art for accurate fraud through a self-consistent hierarchal data detection in dynamic environments pairs model coupled with adaptive technologies. self-learning outlier analytics with a model These techniques are an important approach that incorporates an adaptive feedback to address the industrys need to connect layer. While the outlier analytics adjusts fraud detection across currently siloed fraud detection to changes in the typical decision areas to halt the growing threat of behavior of peer groups, the adaptive layer cross-channel fraud.114
  • 110. Outlier detection across services, products and channels Figure 3The best fraud protection for dynamic markets Figure 4 115
  • 111. References 1. TowerGroup Survey of Credit Card Issuers and Consumer Lenders: Connected Decision Making for Collections, Risk and Fraud Management in Turbulent Times, Theodore Iacobuzio, April 2008, page 17.116
  • 112. Analytics in Financial Services15 Scott Zoldi Alexei BetinProductizing Analytic Innovation: Senior Director, Senior Manager, Analytic Science, Analytic Science,The Quest for Quality, Standardization FICO FICOand Technology GovernanceThe use of predictive analytics is becoming ubiquitous within modern financial IT solutions.However, with the increased complexity of analytic offerings comes the need for astandardized process and build methodology across all analytic model development.The software engineering profession, facing nearly identical constraints, has createdgovernance methodologies that are re-usable for analytic development. As such, the questof the analytics manager is two-fold:1. Apply these effective methodologies to standardize analytic development. This will yield dramatic improvements in the quality and consistency of an analytic teams output, while enabling the flexibility needed for analytic innovation.2. Have analytic technology governance in place to ensure that analytic innovations comply with analytic development standards and constraints. analytics within the software. Analytic model Introduction development across all applications needs to follow a similarly well-defined developmentAs predictive analytics becomes a ubiquitous methodology and process.critical component of modern financial ITsolutions rather than a state-of-the-art add-on Customers of analytics, particularly topfeature, analytic development managers are faced financial institutions, are increasingly aware ofwith new challenges—ones that go beyond these complexities and possible implications.concerns with data quality, choice of training These customers require more transparency thattechniques, latest analytic research, and model an analytic team is utilizing a standard/performance. governance with regard to model-building methodology, technology and process.Mission-critical business applications operateunder many constraints, including stability in This article describes potential solutions andtime, robustness under unexpected conditions, focuses on the practical aspects necessary toand scalability to handle the actual business transform an analytic product organizationsinformation flow. These constraints apply methodology to enable successful delivery,equally to application software as well as to deployment, and operation of hundreds of
  • 113. predictive models annually. The model n Iterative approach to the model development standards and governance development cycle outlined in this article are essential components within a larger set of analytic Tools Governance requirements for the business, which may include legal and regulatory requirements. Development of robust and high Building upon Software quality analytics includes the need for Experience strict adherence to a standardized, approved, and tested tool set. Well-tested Software engineering has matured over time tools are paramount to the correct by developing processes and methodologies development of the model and confident that help ensure consistency and quality of a use of the analytic model in business software release. These methodologies, which applications. ensure that software projects are delivered For example, a sampling script that on time and within budget, started being inadvertently changes the training data set developed as early as the 1960s. Since then, by excluding an important component of they have continuously been improved to the population makes the final deployed meet the demands of the growing size and model not representative of the production complexity of the software products. These environment. methodologies have played a pivotal role in the success of todays software industry. While one can rely on established third- party packages—such as R, SAS, SPSS, or The software development methodology has FICO™ Model Builder—for most typical changed from structured programming to statistical calculations, in practice, there is a object-oriented design and development to large amount of custom code written by the service-oriented architectures and cloud analytics development team. This is needed to computing. The process approaches evolved implement specific custom adaptations of from traditional waterfall to a variety of cutting-edge algorithms, perform ad-hoc data iterative approaches—such as Rational Unified transformations to prepare the data for Process, Extreme Programming, and Agile model development, implement specific Unified Process—developed in recent decades. intellectual property algorithms, and Analytic development can benefit from combine individual calculations into adapting many aspects of software automated modeling steps. development, such as: Without a centralized process, ad-hoc tool n Use of version control and coding standards development results in accumulating too many tools developed by different scientists, n Configuration and release management having multiple versions of the same tool n Development of test cases and automated (sometimes with subtle but important test suites functionality differences), inconsistent documentation and interfaces, and duplicate n modularization and sharing Code or redundant functionality. Further, some (for example, variable libraries) functionality affecting the resulting model n prototyping and integration Early quality and performance may be based on (for example, development of skeleton arbitrary individual choices and not fully models for deployment validation) reviewed, tested and validated.118
  • 114. The solution is to have a well-defined tools creation of a shared variable library.governance process that determines all A standard model code structure andaspects of developing, testing, modifying, organization should exist to facilitatedocumenting, and distributing the common sharing, while leaving the possibility ofanalytics toolset. This ensures consistency customizing variables for an individualacross many scientists developing analytic model—for example, by introducingsolutions. The process often includes coding country and region specific functions andguidelines and procedures for using version variable definitions.control, nightly builds and configuration Code sharing is essential for achievingmanagement. It also covers organizing consistency in how a model handles typicaland documenting various third-partycomponents (compilers, utilities, statistics scenarios and use cases. Response generationpackages, SDKs), and unifying the versions functions, common data fixes, modeland locations of the software across different upgrade logic, error handling and erroranalytic sites and servers. codes can all be implemented in a shared codebase utilized by individual models.While many tools still originate from an ad- The shared code library is accompanied by itshoc solution to a modeling problem, under own automated test suite that invokesthe tools governance process, any discovered common test cases, making it possible togap or deficiency in functionality is brought validate the shared code before it becomesto the Technology Governance Board to part of any model build.consider enhancement of an existing tool orcreation of new standard tools. In analytics development, a shared variable library stands out from other shared code since it encapsulates the accumulated Model Standardization knowledge and Intellectual Property (IP) of the analytics team. Variables are theEven with appropriate tools, it is not possible quintessence of the model. Hence, it isto ensure high model quality if models are paramount for variable definitions to bebeing developed without some form of shared in a version control repository, wherecommon standard and code sharing. Model variables are tested for correctness of thestandardization is necessary to facilitate the definition and leverage consistent variablepropagation of improved or corrected definition standards. Variables are constantlyvariable definitions and bug fixes, and to reviewed for improvement and, through aensure compliance with the latest data common model standard, can be easily madespecifications and consistency of the model available to all models.external interface (for example, reason and Variable standardization and versionerror codes are consistent across models). control also allow for proper inventory ofModel standardization also promotes variables. This prevents redundancy ofcollaboration between project teams variable definitions for those that are closelywho contribute to the standard. This related, and also allows for specific featurereduces duplication of efforts and thereby detector development where variable setsleaves more time for analytic innovation may be smaller. Variables are logicallyprojects. classified based on the types of behaviors to which they respond.The most important aspects of modelstandardization are model code sharing and For example, given a well-organized variables 119
  • 115. library, one can easily determine the variable no-go decisions on utilization of the sets available for cross-border fraud detection modeling data in the build. The report is and, if necessary, focus specifically on new also used on a recurring basis as a tool to variable definitions/improvements. Any tactically improve data contributions improvements can then be available to all where strict adherence with the data future builds, not just to a specific modeling specification/standards is lacking. project where the variable innovation work n Data Preparation – Once high quality may have occurred. data is collected, the data continues to be analyzed from the model build Modeling Process perspective. At this stage, more subtle Standardization issues (such as intersection between various data streams and temporal shifts For an analytics development organization in distributions) are reviewed with the that delivers hundreds of models every year, it analytic manager. Additional sampling is essential to follow a standard methodology and pre-processing is usually performed and process. Each individual model build based on quality and composition of data follows this process, which defines the from different sources (for example, essential milestones and subsequent checks different institutions or sub-systems). and balances that need to be in place throughout the model building process. n Definition and Generation – Variables This involves writing code to calculate The process typically starts with Requirements the variables, striving to provide Definition. Every model may have its own computationally efficient and requirements for the types of input and specific statistically robust (to errors, outliers, and detection targets. Also included would be temporal shifts) definitions. Attention supported versions of the data specification, must be paid to potential legal and and any customer and environment-specific regulatory constraints—such as the Fair requirements and constraints. These can be Credit Reporting Act or non- captured in an analytics tracking document, discrimination laws that may limit or which defines both the requirements and the prohibit the use of certain input fields. deliverables for the model build. Variables that are part of the model The requirements and data availability standard have fully automated test suites determine the choice of the model and are approved for use in models. New architecture and modeling technologies. variables are reviewed by the Technology These choices can be refined as the process Governance Board for approval in a build, goes iteratively through the following typical but also for future inclusion in the model steps: standard codebase. n Data Analysis – This task includes n Selection – This is usually done Variable ensuring the proper acquisition and data using a variety of techniques—such as quality of contributed data for a model correlations, regressions, mutual build. This should consist of a standard information, and sensitivity analyses. Data Quality Report (DQR) that Specific review thresholds related to summarizes the data issues, and contains sensitivity are used to point to any the basic statistics for all the input fields problematic variables that could indicate and sets of automated alerts to data issues. issues—such as target leaks or spurious These alerts need to be reviewed for go/ correlations.120
  • 116. n Training – This presentsModel deployable model code. Many importantchallenges, since many training steps (such as model response calibrationtechnologies, such as neural networks, and input capping, continuity andare inherently complex and involve functional testing) are performed prior todifficulties in explaining the results, delivery of the final model.parameter optimization and tuning,possibility of overtraining, and Quality Assurance Standardinstability to initial model weightsvalues. Standard stopping criteria andmultiple validation data sets are used and Analytic models should be treated like areviewed to ensure robustness. software release, and subject to complete functional testing as it goes through thenEvaluation and Validation – These software integration and Quality Assurancetechniques are not straightforward since (QA) process.they usually involve multiple populationsegments and data from multiple sources. Since the final model is essentially executableFinal validation involves using a hold-out code, bugs may go undetected on modelingdataset that has not been used in any way data, but present themselves when the modelduring training, as well as techniques encounters the real production data andsuch as out-of-sample testing, cross and environment. Similar to software testing,leave-one-out validation. functional test cases are developed and incorporated into an automated test suite.nCreation – This involves creationModel These often involve using synthetic dataof a deployable model, usually by means (including erroneous data) designed toof automated tools that convert the raw expose all potential execution paths andoutput of a training algorithm into error-handling scenarios in the model code. Analytics modeling process and standards for model development Figure 1 Data Analysis Data Preparation Tools Variables Governance Definition and Generation Model Variable Standardization Selection Technology Governance Training Evaluation and Validation Quality Assurance Standard Model Creation 121
  • 117. The tests include model upgrade and score high-level project definition, often continuity tests to ensure that the model formally documented in an Analytic upgrade will not adversely impact the Tracking Document where stakeholders operational environment. from different departments agree on the project requirements, scope and budget. To facilitate seamless model integration, a Model Delivery Standard is utilized. This n Elaboration Phase – In this phase, key standard defines and documents all the assumptions are validated. The interfaces between the software and the development work focuses on the areas of model, as well as the accompanying test data high uncertainty and risk through and documentation. validating the design. This phase involves finalizing the design, obtaining key Analytic Innovation and experimental results and code proto- Technology Governance typing. Any significant change in design/ approach is updated in the Analytic To address a maturing analytic development Tracking Document and requires re- methodology, an Analytics Innovation and approval before moving to the Technology Governance Process must exist to Construction Phase. ensure that only fully validated and approved n Construction Phase – This phase focuses technologies are used in mission-critical and on developing the final version of the productized analytics. This process takes analytic technology based on the innovations from concept to productization, and standardizes the technology prior to results from the Elaboration Phase. adoption in individual models. The production code is written, the necessary modeling steps are identified, The basic principles of a Unified Process and new test cases are developed for (such as RUP or Agile UP) developed in the technology. This analytic technology software engineering can be successfully is then ready for final evaluation adopted in analytic development, and incorporation into the model especially with regard to managing standard. innovation projects. n Phase – The delivery from the Transition A Unified Process defines four project construction phase is reviewed by all life-cycle phases. It emphasizes iterative stakeholders with respect to the value development where different project proposition of incorporating the analytic activities—such as requirements definition, technology into the standard product design, coding, and testing—are mixed in suite. If the innovation meets approvals, a different proportion at each project phase. roadmap item is established to determine n Phase – This phase is based Inception when work will be performed to on an initial idea that often results from incorporate the innovation into the a new requirement, a discussion with model standard. Once in the model the customer, an observation during standard, the analytic innovation is model development, or a research available for use across all analytic article describing a new algorithm. development, the modeling process is The inception phase involves establishing modified accordingly, test cases are an initial set of requirements for integrated into the test suite, and any the innovation, initial design, and necessary new tools are added to the experimentation. This results in a Standard Toolset.122
  • 118. Phases and activities in a typical innovation project Figure 2 Governance for incorporation of formally Conclusion approved, tested, and validated analytic innovations into the Model Standard.Analytic Development Methodologies arematuring. Mission-critical business To meet these requirements, leading analyticsapplications call for high performance teams are exploiting Software Developmentanalytic models, but in equal measure, Methodology practices and adapting themrobust, stable, and predictable execution for Analytic Development. Clients arein their production environment. This b e n e f i t i n g f ro m t h e s e i m p ro v e drequires that the Analytic Development methodologies through consistency in theMethodology adopt Tool Governance/ development process, consistency of modelVersion Control, Model Standard, code, and stability of the models that driveModel Process Standard, and Technology their business decisions. 123
  • 119. Analytics in Financial Services16Analytics in Retail Banking: Anjani Kumar Senior Project Manager, Raghavendra Shenoy Associate Consultant,Why and How? Infosys Technologies Limited Infosys Technologies LimitedIn a service-based economy, companies strive to derive revenue by creating and nurturinglong-term relationships with clients. A case in point is retail banking, where customervalue is of utmost importance. In todays hyper-competitive environment, banks areaggressively leveraging their customer base to engage in revenue driving activities suchas cross-selling and up-selling. To be successful, it is imperative for banks to embracethe power of analytics to gain insights and appropriately evaluate risks andopportunities—enabling more effective decision-making in the quest to enhance walletshare. This article examines the various applications of analytics in retail bankingand provides pointers for analytics implementations. In this context, it is imperative for banks to Introduction offer an interactive and consistent online banking experience coupled with high-qualityIn retail banking, where customer value is at branch banking service. To do so, and tothe core of operations, creating and nurturing achieve faster time-to-market, it is crucial forlong-term relationships with the customer is banks to anticipate customer expectations wellthe key to maximizing wallet share. Advances in technology and the emergence of multipleservice channels has resulted in customers Banks also need to ensure that their existingusing personal computers and smart phones to customers remain satisfied with service qualityaccess banking services. Hyper-competition, and offerings. The cost of customer acquisitionloss of “personal touch” and the use of the is much higher than the cost of customerinternet as an effective channel, has resulted retention. Considering the above, it is essentialin reduced stickiness and switching costs of for banks to effectively use analytics to enhancecustomers, denting bank profitability. customer value and maximize wallet share.
  • 120. Customer acquisition cost is much higher than Figure 1 customer retention cost Reten tion Acqu isiti on Bank ’s C o st Bank Challenges that make channels and an increasingly diffused Analytics an Imperative customer base. Finally, customers are increasingly demanding personalized, In the absence of a robust analytics customized and real-time product offerings, solution, many banks grapple with requiring banks to manage many products— numerous challenges, ranging from the a significant IT challenge. evolving nature of their competitive Strategic: In the new world of banking, environment to regulation and disruptive old strategies (i.e. blanket cross-selling) technologies. A number of these are often prove to be costly and ineffective. highlighted below. Many banks have been criticized for Environment: The financial services focusing on customer acquisition rather operating environment is becoming than retention. Changing strategies increasingly challenging—acquisitions mid-stream has proved to be difficult— are not yielding desired results, organic many banks do not have the infrastructure growth is becoming difficult, new branch required to identify what they want, nor growth is proving costly, and locations the data management capabilities to figure have been rendered irrelevant by online out where to start. Segmentation strategies players. Furthermore, there is increased are difficult to develop in an environment pressure to shrink the product development p l a g u e d b y s i l o e d i n f r a s t r u c t u re . lifecycle. This limits a banks ability to effectively cross-sell and up-sell. Customers: Customer loyalty can no longer be taken for granted. In spite of Tactical: Banks have failed in making CRM technology investments, banks are “emotive” contacts with the customer, unable to maintain a close customer despite having success in “logical” aspects relationship; in fact, many customers like correcting errors. Many banks rely on dont feel valued by their bank. inefficient blanket outbound marketing, Compounding this bank-customer and are unable to leverage focused and disconnect, is the emergence of multiple segmented marketing. In the online channel,126
  • 121. A banker’s challenges Figure 2 Environment Regulatory Customers Banker’s Challenges Competition Strategic Technology Tacticalbanks are unable to provide satisfactory Competition: In a hyper-competitiveanswers to many customers queries. These environment, the banks traditional powerfactors compound one another, making of market share, pulling in deposits,it difficult for banks to sell to their cross-selling, and pricing has waned.established customers. Non-banks are grabbing market share (for example, in credit card and depositTechnology: Banks have immense amounts substitutes). Money portals (for example,of customer data, but many lack the PayPal) which have real-time transactioninfrastructure to predict customer behavior analytical capabilities, pose a tremendousand provide appropriate responses. They challenge to banks. Competition has reducedstruggle to pull in disparate data from various switching costs and lack the capability to track acustomer profile across channels. Regulatory: Customer privacy regulations (for example, spam rules, opt-out programs)Many are stuck with legacy systems—often built have made advertising difficult. Banksto support siloed business processes producing in many countries are also seeing newfragmented data. Such systems have isolated regulatory directives (for example, inmarketing and sales, leading to sub-optimal lending). In the absence of a robustcustomer service. Traditionally, a banks analytics solution, banks are unable totechnology investments have been only in optimally judge their risks andareas of problem and escalation management, exposure—while meeting the increasedwhere customer hygiene is given priority pressure to lend. Business analytics toolsand innovation takes a back seat. Although can help in creating refined customer riskessential, this approach is reactionary. profiles. 127
  • 122. Analytics–Key to (see Figure 3). Analytics can cover a gamut Differentiation of banking functions (see Figure 4). Campaign Management: Using analytics, Banks spend a lot of energy in studying the right offer for multiple product customer data—with a goal of understanding campaigns can be determined. A campaign drivers of customer attrition, repositioning optimizer helps score competing product offerings and using targeted marketing to campaigns against each other, helping to differentiate their services and help improve limit the number of campaigns per customer retention. For a bank to meet its customer to those with the best scores. objectives, a robust analytics solution— Predictive response scoring, channel selection incorporating defined metrics that provide for campaigns, campaign performance a unified view of customers, across lines of monitoring, and cost-benefit analysis are businesses and channels—is crucial. Good all tools that can be leveraged for effective analytics provides banks with many benefits lead generation. Benefits of Analytics Figure 3 Improved Customer Acquisition Reduced Improved Customer Profitability Attrition Enhanced Maximize Customer Interest Value Revenue Benefits of Analytics Improved Decision Maximize Making Fee Revenue Support Enhanced Increased Credit Risk Cross-Sell Management Revenues Reduced Non-Credit Losses128
  • 123. Banking functions where analytics are useful Figure 4 Customer Campaign Profitability & Marketing Management Lifetime Value Analysis Attrition and Service Transaction Loyalty Request Corporate Behavior Management Analysis Function Analysis Cross-sell Costing Predictive and Product Analysis Modeling Holding AnalysisMarketing: Blanket and unsolicited Attrition and Loyalty Management: Usingmarketing interactions lead to customer analytics, segment-wise customer serviceannoyance and wasted marketing satisfaction levels can be gauged. Targetedexpenditure. Analytics help reduce the offerings—of only relevant products andn u m b e r o f u n t a rg e t e d o u t b o u n d personalized communication, usingmarketing contacts with customers. preferred customer channels—will help gainInbound channels (like the internet) can customer more effectively utilized—messagingand offerings tailored to individual Utilizing analytics along with predictivecustomer needs. Cross-channel analytics models to analyze past usage, customertools help identify the most appropriate service logs, and spending patterns, banksinbound channel. Using thousands of can establish early warning systems thatvariables in an automated analytics model indicate customer attrition. This helps deviseto calculate product propensity, banks appropriate retention strategies.can identify the best product/ price/time/ customer/ channel match, and tailor Attrition scores can be computed to predictofferings appropriately. Use of event- attrition probability at the account andtrigger engines can help alert the bank of customer levels. Profitable customers atmarketing opportunities proactively (for risk of leaving (for example, at a mortgageexample, insurance when a person moves). tenure end) can be identified. Offers to encourage such customers to renew theirCustomer Profitability and LifetimeValue Analysis: Using analytics, products relationship with the bank can be made,can be priced according to a customers based upon past experiences and thelikely future value. The long-term customers profile. Analytics also help banksprofitability of a customer can be gauged by design effective loyalty programs byanalyzing different cost and revenue providing insights on customer loyaltycomponents across products. parameters. 129
  • 124. Service Request Analysis: Analytics can cost channels. Cost containment through help monitor customer satisfaction on effective credit risk analysis is also possible. service quality, by providing insights on Predictive Modeling: Analytics can be evolving customer needs and satisfaction leveraged to analyze past customer behavior levels. Customer satisfaction levels on non- and, thereby, predict future behavior. financial interactions made with the bank can Sometimes, a rapid increase in customer base, also be monitored using analytics. through organic or inorganic growth, creates Corporate Function: The sales decision- huge challenges for banks attempting to making process can be strengthened by know their customers better. Predictive linking analytics to fraud and money- analytics can be of great help here. As laundering detection and credit/ risk scoring. customers mature, they shift into new Analytics can also help to get the marketing segments. Analytics provide insights into department involved in product configuration, evolving customer needs during various life pricing and placement decisions. stages, allowing the bank to evolve along with Another potential function is developing the customer. multi-dimensional views on aggregate, Cross-Sell and Product-holding Analysis: segment-wise and trend insights to help Cross-selling and up-selling are crucial to banks identify focus areas for maximum ensure higher wallet share. Analytics can improvement. Location-specific insights be used to check spending patterns and can also be attained and strategies defined other customer behavior, thereby strategizing (for example, focus on customer acquisition for cross/up-sell. Using a prospects detailed in one market and on retention in another). profile and behavioral information, banks Market behavior can be understood better can pre-select high product propensity and responses built. Expeditious and customers and make targeted offerings on meaningful branch goals can be set by their preferred channels. Proper customer intelligent segmentation. segment analysis and understanding of Transaction Behavior Analysis: Analytics optimal profit-to-risk mix, using analytics, can help financial institutions analyze can help banks sell more to established transactional behavior aspects (for example, customers. Perhaps most importantly, recency, frequency, and the monetary value of analytics can be used as a tool to achieve the a customers transaction and profile). They goal of a connecting hub, meeting all also help reveal channel preferences, and financial needs of the customer. usage for specific products and transaction patterns across customer segments. Analytics Implementation Approach Cost Analysis: Analytics help compute operating costs per activity type. The cost To enable banks to fully benefit from can be referred for various combinations analytics, a structured approach for of channel, product and customer segments. implementation is crucial. Many banks Customers using live channels (for example, follow a three-phased approach. In the branch, call centers) cost much more to serve first phase, banks master customer data. than customers using self-service channels. In the second phase, banks attempt to Analytics are very useful in determining gain insights through automated analysis channel profitability; banks can then design (for example, profitability and cost) of strategies for migrating customers to low existing data. In the final phase, banks130
  • 125. Analytics implementation – Three-phased approach Figure 5 Hindsight Insight Foresight Data Knowledge Wisdom Data Knowledge Wisdom Customer ? Product ? Customer ? Information Profitability Value-oriented File Customer ? strategies / Profitability tacticsProposed analytics infrastructure Figure 6 Analytics Infrastructure Intelligence Business Intelligence Cross-Sell Customer Analytics Attrition Transaction Service Campaign and Product Profitability and Loyalty Behavior Request Management Holding and Lifetime Analysis Analysis Analysis Analysis ValueConsolidation Data ETL and Date Cleanup Data Analytics Warehouse DatastoreSources Data External Customer Account Transaction Data Service Data Data Data Request Data 131
  • 126. attempt to build rules and predictive models Retail Banking Analytics- based upon insights gained in phase two Things to Ponder (refer to Figure 5). In the first phase, banks typically dont Before undertaking an implementation, falter. The second phase is where things financial institutions should consider the become challenging. In this phase, banks following. often lack coherent strategies and attempt Basics: Before delving into analytics half-hearted analysis, top-down profitability set-up, banks must have a core banking analysis (rather than bottom-up), and/ or solution in place. The basic needs of the a siloed, unit-level undertaking. bank should be met first. If systems are In the third phase, planning and preparation inefficient (for example, the account opening is critical. Successful completion of process is too long), they must be addressed phase three requires a bank to have built first. Pulling data from different channels an organizational infrastructure for and sources should be possible and relatively harnessing analytics, customer segmentation, efficient. Finally, banks having relatively and actual/ potential life-time value crude metrics must plan for more computation. sophisticated ones to be used with analytics. Effective analytics implementation Strategies: Fail to plan and plan to fail. Banks necessitates a flawless transition of business must have a strategy on the information they strategy into analytics architecture strategy. want to collect. The right type of data All applicable analytics architectural stacks, collection, not just basic banking transactions, including Base Infrastructure, Data, is crucial. Where complexities are immense, Discovery/ Integration, Analytics Applications, banks must start small and then scale up. Performance Management, Reporting, and Furthermore, for an effective customer- Delivery must be thought through and planned profitability analysis, a bottom-up approach for implementation. Refer to Figure 6 for is recommended (for example, building from a proposed analytics infrastructure. the customer account level). Effective analytics Figure 7 Real Basics Time Experts Strategy Social Cross Media Functional Effort Effective Analytics132
  • 127. Real-time Analytics: At times, a banks Experts: Banks must provide expert employeesanalytics dont lead to real-time promotion of the discretion to augment the analytics outputproducts or pricing. This is because many with their own knowledge, where appropriate.banks implement analytics with bought-in Sales agents should have the final word. Bankstools that have simple propensity models using analytics must train and trust their agentsinstead of real-time event triggers and to make the final decisions.personalized pricing. Building real-time Social Media: Banks should leverage socialpredictive analytics capabilities will give media and Web 2.0 features (for example,banks a “bigger bang for their buck”. Take the social networks, wikis, blogs, podcasts). Theyexample of a daily data scan to identify should align these with analyticscustomers fitting the cross-over or retention infrastructure to gain useful customerrisk pattern. Here, real-time analytics can insights. Banks can consider partnering withenable automatic alerts to customer-facing online social networks.personnel (for example, account managers)for timely action. ConclusionCross-Functional Effort: Effectiveimplementation of analytics is a cross- Understanding and utilizing the power offunctional effort. Close collaboration analytics has become imperative for retailamongst all concerned lines-of-business banks. Potential applications of an effectiveand technology leaders is crucial in defining analytics program are nearly limitless. Whenholistic analytics strategies and eliminatingsilos. To build good, predictive analytics, implementing, banks must follow a structureddeveloping robust business rules is crucial. approach and ensure that specific needs areFor this, all concerned business units of the kept in mind, proper stakeholders are involved,bank need to contribute the typical patterns and a long-term strategy is developed. Thoseof their businesses. As with any major that do, will achieve a significant competitiveinitiative, senior management support and a advantage—selling the right products to thecollaborative culture is crucial. right customers at the right time. 133
  • 128. Analytics in Financial Services17Business Analytics Guruprasad Rao Senior Industry Principal Harpreet Arora Client Engagement K N Rao Principal Tech Architect, & Head of IT Consulting, Manager, Head, BI Consulting,in the Wealth Management Architecture & Innovation Banking and Capital Infosys Technologies Enterprise Solutions Markets, LimitedSpace Infosys Technologies Infosys Technologies Limited LimitedTodays difficult investing environment has (if nothing else) driven one thing—anincreasingly demanding breed of investors. From managers of institutional investments, tohigh net-worth individuals, to trusts, investors want an increased breadth and depth ofinformation. To effectively respond to these demands, wealth management firms must investin improved information management and analytics. This article provides a roadmap forfirms to harness the power of information and satisfy the new breed of investor. corporations, and trusts. Catering to the needs Introduction of these clients is a team of advisors, researchers, client relationship managers, portfolioWealth Management is undergoing a major managers, traders and brokers, that work to helpfacelift these days. The golden era of eternal clients make investments across a swath ofincreases in investment values has passed. The products—equities, debentures, mutual funds,credit crunch, recession, regulations, and futures, works of art, commodities, fixedcontinued global financial instability have deposits and ETFs.forced consumers and wealth managers alike toerr on the side of caution. To achieve the necessary portfolio growth, firms are increasingly learning to harness the power ofDespite a rapidly evolving competitivelandscape, wealth management firms need to their information—often through analytics.adapt to economic and market developments, However, to accomplish this, they must firstand continue to focus on their clients understand their data and structure itportfolios. appropriately.For most institutions, these clients take theform of high net-worth individuals (HNIs),
  • 129. Wealth management lifecycle Figure 1 TRANSFER ? Giving Charitable ?Legacy Planning ?Wealth Stewardship UTILIZATION TRANSFER ? Income Planning Retirement ?Inflation Protection ? Care Planning Long Term UTILIZATION PRESERVATION ?Risk Management ?Tax Mitigation PRESERVATION PROTECTION ?Business Succession Planning PROTECTION GROWTH ?Planning Insurance ?Asset Protection ?Portfolio Risk Mitigation Wealth Management Lifecycle GROWTH ?Asset Allocation ?College Planning ? Planning Retirement Managing a Wealth of periodic basis. Table 1 highlights reports that Information are often prepared. Complex Information Requirements Disparate Information Depending on the client relationship, The investment advisory teams work on investment decisions are made either on a different sets of applications as they deal with discretionary basis or a non-discretionary a variety of investment products. Clients may basis. Wealth management organizations not be interested in all the investment earn their revenue based on a fees and avenues offered by a wealth management returns basis, so it is imperative for them unit, hence their details may be scattered to track the value of these investments and across various trading applications. At times, align them to the market on a periodic basis. this results in storing customer details Apart from this, the wealth management differently across disparate applications. firm is also interested in understanding Some applications need to store investment which investment advisor is generating details for a short time only, while others more income, which portfolio managers need to store the details for a very long mutual funds are doing well, and which time—primarily due to regulatory demands. researcher is performing research with a high-value-add that can be leveraged for A typical wealth management team needs to greater returns on the investment. prepare a list of reports for clients as well as for the bank (if not independent) on a Wealth management firms are also interested1366
  • 130. A wide array of wealth management reports Table 1 Report Type Description Performance Performance of the various components of the clients investments, Attribution across different industries and instruments Risk Market Risk statement to the customers on their portfolio Cash Flow Cash Flow information to the customers Fee Paid Fees collected and invoiced to the customers Holdings Exhaustive list of all the holdings in different forms Valuation Mark to Market valuation of the Investments /Assets Benchmark Comparative statements: Plan Vs Actuals Comparison Corporate Action Dividends due & Corporate Actions due Reconciliation of the expected cash flow from sale / purchase of assets, Reconciliation dividend allotments with the physical / demat list of certificates Physical Certificates List of Assets in Physical Certificate form List Cash Management Reconciliation, Overnight balances, Overnight calls Securities Reconciliation statement Reconciliationin understanding the risk appetite of their earnings/ losses and other fees and costsclients and the nature of their investment involved in various transactions.portfolio—both critical in suggesting a To get started, data must be collected fromsuitable investment strategy. trading systems and other record storage systems and assembled to reflect individual Information Management Framework client portfolios. A complete list of master data (including the customer master dataAll this calls for effective information and trading reference data) also needs to bemanagement and dissemination to all maintained. This master data hub can alsoparties concerned. The Information act as a golden source for master data andManagement Framework (IMF) should meta master data for the entire unit/enable the establishment of a single version organization. The proper communicationof the “truth”, and enable the viewing of and exchange of meaningful informationthe complete investment portfolio segregated across various individuals, departments,by clients/ sectors/ investment arenas/ and stakeholders is critical.profiles. Such a framework should also Most importantly, for the IMF to function,associate the influencers and decision-makers the source systems data must be arranged in awithin the wealth management organization meaningful manner, with a provision to storeto specific actions/ decisions. Finally, the detailed transactions and other details onframework should effectively reflect the clients, products and industries. Provisions 137 7
  • 131. Information management framework components Figure 2 SOLUTIONS PERSPECTIVES FRAMEWORKS ERP, CRM, Banking, HRMS, Legacy BUSINESS INTEGRATION TRANSACTION SYSTEMS PLANNING & DATA WAREHOUSE EAI, ESB, SOA BUDGETING BUSINESS DECISIONS ANALYTICAL WORKFLOW PLATFORMS INDUSTRIES ANALYTICS should be available to look at aggregated provides a single version of the truth made data and then drill down to the details available in a prioritized and easily when required. understandable format. In such a manner, the IMF unites the entire The set of KPIs and Metrics can easily be wealth management unit for information- divided into the following subject areas: sharing through master/ reference data n · Assets Growth standardization. It also enables the unit to bring out the single version of truth, as it n Growth/ Churn · Clients integrates the source systems extracted data n · Assets Quality/ Products at a target database in a single schema. n · Credit Risk Performance Management n · Operations Risk Framework n · Other Financial Metrics A Performance Management Framework n · Employee Productivity (Trader, (PMF) can be defined on top of the Investment Advisor, Client Relationship established IMF to monitor the Manager) organizations health on a number of n · Employee Effectiveness (Investment aspects. This is accomplished through the Advisor) use of a well-defined set of KPIs and Metrics. Essentially, the PMF on top of the IMF n Effectiveness · Process13868
  • 132. Logical architecture of information management framework Figure 3 Enterprise Data Sources (Multiple, Disparate) 3rd Party XML/ Applications Databases Flat Files Interfaces Unstructured Information Processing & Integration Information Information Information Acquisition Consolidation Transformation Information Organization Information Enrichment Verification & Validation Graphical Analysis Information Presentation & Sharing Information Analysis & Presentation Delivery Customer Content Partners Channels Business Analytics DimensionsSome of the applicable KPIs/ Metrics under create an interactive environment wherethe above-mentioned subject areas are clients carry out a variety of simulations—expanded upon in Table 2 on the next page. hypothetically redistributing their investments across different asset classes Business Analytics and geographies. BA, by the way of a Framework Global Balanced Asset Allocation System (GBAAS), can play this role.Business Analytics (BA) play an even The important inputs required forgreater role in increasing the wealth of the BA environment to function are asthe clients, by providing critical insights follows:into the markets, geographies andeconomic trends. BA enables this by n data in dimensional hierarchy, · Masteranalyzing the data that is already along with the customer investmentgathered through the IMF and perfected by profile and risk appetitethe PMF. n data arranged as per geographies, · ResearchFor example, by using analytics, a firm can industries and focused organizations 139 7 9
  • 133. Applicable KPIs and metrics Table 2 SI No. Subject Area KPI Parameters 1 Assets Growth ?of business growth across category, with existing Volume clients/new clients ? business growth across category Speed of ? order conversion ratio, with discretionary/ Advise to non-discretionary? ? growth across categories Net new 2 Clients Growth / ? of new clients added/ churned Number Churn ?categories exited /entered by new/ existing clients Product ? growth in clients Net new Client satisfaction ? Client interactions ? Cross-selling results ? 3 Assets Quality / Assets category performance ? Products ?assets at risk/ problem across categories Clients Client portfolio performance ? 4 Credit Risk ?actual credit risk vs preferred risk appetite Clients Asset portfolios actual credit risk vs allowed risk ? 5 Operations Risk ? Concentration Risk Industry Compliance Risk ? ? Risk Statutory Documentation Risk ? ? Risk Settlement 6 Other Financial Metrics ? return on economic capital Risk adjusted ? risk adjusted assets Return on ? Liquidity 7 Employee Productivity ? Client on-boarding efficiency ? trades executed per hour Number of ? settlements done per hour Number of ? Mark to Market (MtM) done per hour Number of 8 Employee Effectiveness ? discretionary vs. non-discretionary agreements Number of ? quantity by category Trades/ sales ? Target Achievement 9 Process Effectiveness ? Staff at different operations Number of ? SLA adherence ? mistakes Number of n · Projections on the research data for about n to drill down from the KPIs to · Ability 10 years dimensional architecture, and then to contributing detailed transactions n data categorized and profiled · Research into different risks n · Most importantly, the ability to model the data for some projections and n · Availability of statistical tools, interactive simulations visualization tools and interactive decision-making data architecture1406
  • 134. Data Integration (DI) Technology Framework ComponentThe technology framework should address all The functionalities expected from the DIthe three frameworks—Information Component are:Management, Performance Management and n · Extraction: Automated data extractionsBusiness Analytics. This forms the from internal and external systems, eitherunderlying foundation for addressing the in a data pull or push method.functionalities. Sometimes, for important data, it isThe important components of the Technology required in a real-time manner.Framework are: n · Validation: This is very important ton· Data Integration (DI) ensure the accuracy and completeness of the extracted files.n· Data Management (DM) For the data integration component to workn· Data Arrangement (DA) efficiently, without interfering with then· Data Exploration and Analysis (DEA) performance of the source systems, the Technology architecture for business analytics Figure 4 Internet Mobile Intranet Stock Market Security (Identity & Access Management) Feeds Portal Server External Analyst Feeds Business Process Management External B2B Internal Risk Allocation Analyst Feeds Simulation Modelling Engine Equity ESB / SOA Research data Market Risk Business Intelligence Directives Enterprise Apps LDAP Intermediary Operational Legacy Data Store Apps Asset Manager Analyst Warehouse Desktop 141 7
  • 135. required data is extracted in as-is manner into hubs “golden sources” of master data for the a staging area, in text files. From here, the data entire organizations use. is loaded into tables for easier handling. Data Exploration & Analysis Data Management (DM) (DEA) Component Component This layer forms the interaction and The functionalities expected from the DM exploration layer for the business users. The component are: functionalities expected from this DEA component are: n Data cleansing is required for · Cleansing: data coming from source systems that is n · Allow external and internal users selective not clean or standard. This step ensures access to the data by defining security that the data is standardized: using one rights. set of rules for the format and meaning. n various exploration means in the · Provide n · Transformation: The transformation form of standard reports, ad hoc reports, rules perform simple to complex KPIs/ Metrics and alerts through calculations and arrange master data in Dashboards. hierarchies. n visualization tools also come · Interactive under this category—enabling users to Data Arrangement (DA) explore the data in their own way to Component identify outliers and trends. The functionalities expected from the DA n · Statistical tools enable the users to component are: perform some future projects and trends nModel: Schema is the most · Data using the historical data. important overall technical component. To be effective, the Technology Framework This dictates the way the data is arranged must have the capability to seamlessly to ensure consistency and accuracy. integrate and orchestrate with back-end Usually, a normalized schema is required applications, legacy applications, and for these activities. enterprise applications. n · Dimensional Data Model: This type of data schema ensures an interactive Conclusion environment for the business users. This also allows the data to be drilled up and Several leading banks in the wealth down along the hierarchies presented in management space are in the process the master data. of making major improvements in their information management capabilities. n · Sometimes data may be arranged in In most cases, these changes are occurring departmental data marts, in a in the statutory area and risk management. combination of dimensional and However, this is only a stepping stone. aggregated techniques, to enable the Wealth management firms can tap into departmental users to exclusively work the potential of business analytics to on them. differentiate and ensure customer Data arrangement also makes the master data satisfaction.1426