Getting Your First Predictive Model Up and Running

Getting your first predictive model up and runningJames Taylor, CEO

Decision Management isA business discipline that builds onexisting enterprise applications to put data to work manage uncertainty increase transparency give the business control ©2009 Decision Management Solutions 2

5 core principles of decisioning Identify, separate and manage decisions Use business rules to define decisions Analytics to make decisions smarter No answer is static Decision-making is a process ©2009 Decision Management Solutions 3

Delivering Decision Management 3 stages to better operational decisions Create a “closed loop” between operations and Design and build analytics to independent measure results and decision processes drive improvement to replace decision Identify the points embedded in decisions (usually operational systems about customers) that are most important to your operational success ©2009 Decision Management Solutions 4

About today’s presenters Anunay Gupta Co-founder and Head of Analytics, Marketelligent 10 years of experience in Consumer Banking, Risk Management and Decision Management at American Express and Citigroup Work with clients to leverage their data for strategic and tactical decisioning

What you should get out of this webinar• English-version overview of what folks mean when they talk about business intelligence, analytics, predictive analytics, models, scorecards, etc…• Some idea of the mathematics and the sophisticated techniques that work behind-the-scenes• More importantly, real-life situations where you can leverage Predictive Analytics to drive profitable growth in your business• In the end, its not rocket science. However it does require specialized skills and expertise for successful build and deployment 6

Agenda • What is Predictive Analytics • Critical Requirements for success • Real life applications  Direct Marketing : Maximizing ROI  Consumer Finance : Whom to sell? What to sell? Which Channel?  Consumer Packaged Goods : Marketing $ Optimization • Summary • Q and A 7

www.puntersgenie.com……….we take as much historical data from racing as we can and try to find the things that are important for predicting the outcome of future races. Once we find those things (in some cases we can be working with tens of thousands of combinations of variables), we then run the models against a test set of races and look at the results. We then look at the races that we predicted correctly and work out what things made that possible for those particular races. This is how we come up with the Bet Index. This information is then fed back into the models to make them better Predictive Modeling …. predict the probability of a horse winning a race 8

What is Predictive Analytics ? “Use historical data to make certain predictions for the future” Hindsight Insight Foresight “What will happen?” “What is happening ?” “Why is it happening ?” “What should happen?”  Typical MIS or BI  Business analysis  Predictive Analytics;  Cognos; Business Objects;  behavior analysis; trends; forecasting; optimization, Hyperion; ProClarity; etc etc etc  Largely backward looking  Gives us insights on what  Uses past behavior to is happening and why predict future outcomes  Referred to by many folks as ‘Analytics’ although it  Game changing is not  Forward-looking 9

Some types of Predictive Analytics Logistic Forecasting; Segmentation; Regression OLS; ARIMA CHAID; CART  Commonly used when the  Used to forecast  Used to bucket or ‘cluster’ objective is to predict a outcomes that are of a like things binary outcome continuous nature  Each member in a cluster  Example: will Customer X  Example: how much will is very similar to another respond or not respond to this Customer Y spend in member in same cluster; my marketing offer the next month? but very different from a  Example: What is the  Example: movement of member in a different chance Customer Y will the S&P 500 index on a cluster dis-enroll in the next 12 weekly basis for the next  Example: Customers in a months 12 weeks particular segment have similar behaviorsARIMA: Autoregressive Integrated Moving AverageCHAID: Chi-squared Automatic Interaction DetectorCART: Classification & Regression TreeOLS: Ordinary Least Squares 10

Critical Requirements for Success Business Objective Data Expertise Culture More data is better; Requires folks that Typically Senior and data from are not only management buy- varied information statisticians; but can in is critical. sources is even also understand Successful better business projects are top- driven Predictive Analytics 11

Business Objective I want to identify which Customers will ‘attrite’ so that I can take some proactive actions All Customers? Or just new Customers??? Attrite today / tomorrow / next month / etc What is attrition to me? No activity for 6 months / 2 months / etcI want to predict which of my high tenure Customers will ‘attrite’ or ‘churn’ in the next 6 months 12

Analytical Framework Business Objective:I want to predict which of my high tenure Customers will ‘attrite’ or ‘churn’ in the next 6 months Past Future-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 Months 1. Historical Customer transaction data Decision Period (mob>12; transactions, interactions) 2. External data (Credit bureaus; demographics; psychographic, macroeconomic; etc) Decision Point Dec09 13

1. Data Collection Identify a suitable time period in the past to collect relevant information Past-25 -24 -23 -22 -21 -20 -19 -18 -17 -16 -15 -14 -13 -12 -11 Months 1. Historical Customer transaction data Decision Period (mob>12; transactions, interactions) • Identify Attritors; label them as 1’s 2. External data • All others labeled as 0’s (Credit bureaus; demographics; psychographic, macroeconomic; etc) Reference Point July08 14

2. Model Build & Deployment Model Raw data Exploratory Data Variable Variable Development & Deployment Analysis Treatment Selection & Sampling Validation Data Preparation  Defining  Missing Value  Stepwise  OLS / Logistic /  Scorecard Over sampling ? dependent Treatment regression CHAID / etc development Reject variable  Variable  Logit Plots  KS  Statistical paper Inferencing  Business sense Transformation  Business Logic  Rank-ordering  Implementation check  Variable capping code  Multi-collinearity  Out-of-time & Flooring Validation  5 – 10 most significant variables Ongoing Model Validation & Maintenance 15

Output of Modeling ProcessEvery Customer has a unique ‘Score’ that captures the essence of what is being modeled.The ‘Score’ is essentially the ‘probability’ of something happening scaled in a pre-defined fashion; having an upper- and an lower-bound Called a ‘Score-card’ For Example:1. Customer #17523 has a score of 769; translating to a 90% probability of ‘churning’ in the next 6 months 2. Household # 845 has a score of 423; translating to a 36% chance of accepting the offer for a magazine if sent a Direct mail Offer 16

Resources & Timelines CRISP-DM Process 20% 25% 15% 5% 25% Business: 30% Data: 40% Modeling: 25% 10% 17

Explaining the benefits Random w/ MIDAS Blaze™ 100% 90% • Save: 25% improvement in marketing efficiency; leading to annual cost% Responders Captured 80% 70% savings of $1.5MM. Same number of Boost 60% Customers acquired 50% Save 40% • Boost: 25% more acquired 30% Customers with a marketing budget 20% of $6MM. 10% 0% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% • Build scenarios and optimize % Mailbase Sell the business impact; not the technical power ! 18

Business Applications • Optimize your Marketing $ Direct Marketing • Maximizing Customer Lifetime Value Consumer Finance • Deepen relationships by cross-sell & up-sell Telecom & Utilities • Retain Profitable Customers • Risk Management & Fraud Healthcare • Collect past-dues faster • Predict Part Failures Manufacturing • Web targeting 19

1. Direct Marketing Cut marketing expenses significantly; while maintaining acquisition volumes Random Mailing Intelligent Mailing Response Rate: 4.5% Response Rate: 6.0% MailedMailed Scorecard Not Mailed : Prospect : Responder Response Scorecards help in identifying Prospects/Customers to target so as to maximize Response rates 20

Final Mailing Strategy 25% improvement in marketing ROI - 6 campaigns of 1MM mailings each; annual cost of $6MM - Random mailing Response rate of 4.5% → 270,000 Responders - Response Model built; assigns each prospect a ‘Response Score’, between 1 and 10 - 9 campaigns of 0.5MM mailings each; annual cost of $4.5MM → 270,000 Responders - 25% improvement in marketing efficiency; leading to annual cost savings of $1.5MM RANDOM MAILINGS TARGETED MAILINGS Response # Cumulative # Cumulative Marginal Cuml # Cumulative Marginal Cuml # Prospects # Responders # Responders Score Prospects Responders Response rate Response rate Responders Response rate Response rate 1 100,000 100,000 4,500 4,500 4.5% 4.5% 9,507 9,507 9.5% 9.5% 2 100,000 200,000 4,500 9,000 4.5% 4.5% 6,761 16,268 6.8% 8.1% 3 100,000 300,000 4,500 13,500 4.5% 4.5% 5,282 21,549 5.3% 7.2%Increasing 4 100,000 400,000 4,500 18,000 4.5% 4.5% 4,437 25,986 4.4% 6.5% Response 5 100,000 500,000 4,500 22,500 4.5% 4.5% 4,014 30,000 4.0% 6.0% 6 100,000 600,000 4,500 27,000 4.5% 4.5% 3,592 33,592 3.6% 5.6% Rates 7 100,000 700,000 4,500 31,500 4.5% 4.5% 3,169 36,761 3.2% 5.3% 8 100,000 800,000 4,500 36,000 4.5% 4.5% 2,958 39,718 3.0% 5.0% 9 100,000 900,000 4,500 40,500 4.5% 4.5% 2,746 42,465 2.7% 4.7% 10 100,000 1,000,000 4,500 45,000 4.5% 4.5% 2,535 45,000 2.5% 4.5% 1,000,000 45,000 4.5% 45,000 4.5% 21

Response Model Performance 10% 9% Modeled 8% 7% Cumulative 6% Response 5% Rates 4% Random 3% 2% 1% 0% 1 2 3 4 5 6 7 8 9 10 Increasing Response Rates If needed, marketing efficiencies can be further increased by targeting high responding prospects 22

2. Consumer FinanceWhat to Sell? To whom? Which Channel Channels Products Customers 23

What is Customer Lifetime Value ? Measuring Customer Lifetime Value CLV is defined as the sum of cumulated Cash-flows – discounted using the Weighted Average Cost of Capital (WACC) – of a Customer over his or her entire lifetime with the Franchise Known from Predict Responseexisting P&L’s Rates Acquisition Monthly Costs Expenses Customer Net Margin Lifetime Value Monthly Accumulated Revenues Margin Customer Lifespan Predict monthly Spend Predict Customer Attrition 24

Eg. Credit Cards CLV(Customer1, product XY, Channel PQ) = f (P&L drivers, discount rate) Customer / Segment Acquisition Cost Acquisition Models: Discount Rate -Product & Channel based -p(Response Score) Total Customers -p(Approval Score) Revenue Models: Purchase Sales, $ -p(Activation) Payment $ -p(Monthly purchase sales) -p(Payment $) Net Credit Losses, $ -p(Attrition) Ending Loan Balances, $ Revenues Expenses Expense Models: Net Income (after taxes) -p(Credit Loss) Terminal ValueModels can be built at Customer- Discounted Net Income level or Segment-level Discounted Terminal Value CLV 25

Eg. Credit Cards Cross-sell Over 80MM Combinations ! 4 Channels Business constraints10 Products Optimize Target Right Product to right 2MM Customers Customer in the right Channel 26

3. Consumer Packaged Goods Optimize marketing spend across channels Marketing-Mix-Optimization Optimize investments across Media so as to maximize Sales Historical data is collected for sales (and/or other KPIs) and Multivariate regression analysis is used to quantify all key Media Marketing activities incremental sales generated$600,000 $600,000$500,000 Past sales $500,000 performance$400,000 $400,000$300,000 $300,000$200,000 $200,000 Incremental sales$100,000 Past TV $100,000 generated by TV activities $0 $0 Week10 Week13 Week16 Week19 Week22 Week25 Week28 Week31 Week34 Week37 Week40 Week43 Week46 Week49 Week52 Week10 Week13 Week16 Week19 Week22 Week25 Week28 Week31 Week34 Week37 Week40 Week43 Week46 Week49 Week52 Week1 Week4 Week7 Week1 Week4 Week7 27

Optimally allocate Media spend to maximize Sales Baseline Sales Magazine Incr. Sales TV Incr. Sales Daily Incr. Sales test Magazine Spend TV Spend Dailies Spend 20 900 18 800 Media Spend, ‘000 SGD 16 700 Volume, ‘000 units 14 600 12 500 10 400 8 300 6 200 4 2 100 0 0 DEC07 DEC08 AUG07 OCT07 AUG08 OCT08 SEP07 SEP08 MAY07 MAY08 MAR07 JUN07 MAR08 JUN08 JUL07 NOV07 JUL08 NOV08 FEB08 FEB07 APR07 APR08 JAN07 JAN08 28

Magazine gives the highest ROI per $ spend Incremental Sales per ‘000 SGD media spend 0.14 0.12 For every $ spend, Magazine gives 6 0.10 times the return of Efficiency 0.08 TV and dailies 0.06 0.04 0.02 - Total Spends Magazine TV Daily 29

Key TakeawaysPredictive Analytics can be a potent weapon in your toolbox With increasing commoditization, it is truly the next differentiator It requires specialized expertise, talent and tools to execute well 30

About Marketelligent anunay.gupta@marketelligent.com www.marketelligent.com 1.201.301.2411 31

Decision Management Solutions Decision Management Solutions can help you Focus on the right decisions Implement a blueprint Define a strategy For assistance, to find out more or if you have questions decisionmanagementsolutions.com/learnmore ©2009 Decision Management Solutions 32

Getting Your First Predictive Model Up and Running

by Decision Management Solutions on Mar 30, 2011

Accessibility

Tags

Upload Details

Usage Rights

Statistics

Getting Your First Predictive Model Up and Running — Presentation Transcript