The present work overviews the application of recommender systems in various financial domains. The relevant literature is investigated based on two directions. First, a domain-based categorization is discussed focusing on those recommendation problems, where the existing literature is significant. Second, the application of various recommendation algorithms and data mining techniques is summarized. The purpose of this paper is to provide a basis for recommender system- and financial experts to work out further scientific contributions in this field.
Recommender Systems meet Finance - A literature review
1. 1
RECOMMENDER SYSTEMS MEET FINANCE:
A LITERATURE REVIEW
David Zibriczky1,2
1ImpressTV
2Budapest University of Technology and Economics
2nd International Workshop on Personalization and Recommender Systems in Financial Services
Bari, Italy – June 16, 2016
2. 2
• Recommender Systems
• Financial products
• Part 1: Domain-based review
› Definition: specific area of finance that can be properly identified, modeled and developed based on its specific properties
› Applications, algorithms, studies, concepts
› Characteristics: 1) heterogeneity, 2) churn, 3) interaction style/volume, 4) preference stability, 5) risk, 6) scrutability
• Part 2: Method-based review
› Standard recommender methods, complementary methods
› Pros/cons in financial domains
Introduction
4. 4
Domain-based review / Loan
Loan: lending money from one entity (individual or organization) to another one with specified conditions
• Goal: Personalized recommendation of loan products
• Application: VITA – Sales support between representatives and customers interested in loan
Microfinance: is a source of financial services for small businesses lacking access to banking and related
services
• Goal: Fair pairing of lender and borrower
• Applications: Peer-to-peer microfinance (Kiva), P2P lending as investment, social network-based lending
Heterogeneity Churn Rate Int./vol. Preference Risk Scrutability
low low/medium explicit/low unstable high high
5. 6
Domain-based review / Insurance
Insurance policy: is a contract between the insurer and the insured (policyholder), in which the insurer takes
obligation to pay compensation for insured if loss caused by perils under the terms of policy
Insurance rider: provision of an insurance policy that that provides additional benefits at additional cost
• Goal: Personalization of conditions based on personal status and goals
• Applications:
› MyINS: A CBR e-Commerce Application for Insurance Policies
› Real-time cloud-based health insurance plan recommender system
› Concept for recommending insurance riders
Heterogeneity Churn Rate Int./vol. Preference Risk Scrutability
low medium explicit/low stable high high
6. 7
Domain-based review / Real estate
Real estate: a property consisting of the land, its natural resources and the buildings on it
• Goal: Personalized recommendation based on multi-criteria for one purchase
• Applications:
› Online real estate search application
› Multi-criteria journey-aware recommender
› Fuzzy Expert System for real estate recommendation
Heterogeneity Churn Rate Int./vol. Preference Risk Scrutability
low high explicit/low
implicit/med.
stable high high
7. 9
Domain-based review / Stock
Stock: type of security that represents ownership in a company and claims on its assets, earnings and dividends
• Goal: Forecasting returns, capturing risk and risk aversion, tradeoff between risk and return
• Applications:
› Non-personalized: Buy/sell signals, forecasting returns
› Personalized: Capturing risk-aversion
- Interactive Graphical User Interface
- Individual transaction or metadata-based solutions
Heterogeneity Churn Rate Int./vol. Preference Risk Scrutability
high low implicit/high unstable very high very high!
8. 10
Domain-based review / Asset allocation and portfolio management
Asset allocation: an investment strategy that attempts to balance risk versus reward by diversification
Portfolio: a weighted composition of finite number financial assets and other investment
• Goal: Maximizing risk-return ratio based on user preference or risk-aversion
• Applications:
› Non-personalized: Modern Portfolio Theory and its extensions, diversification
› Personalized:
- Case-based reasoning
- Fuzzification of preference of users and ontology of portfolio
Heterogeneity Churn Rate Int./vol. Preference Risk Scrutability
high low, high explicit/low stable high high
9. 11
Domain-based review / Other financial domains
Venture capital: a type of private equity that is offered for start-up companies as seed funding
• Goal: Advantageous paring of VC firms and start-up companies
• Characteristics: Risky, high item churn, extremely sparse data, conventional methods are not efficient
Stock fund: a fund that principally invests in stocks
• Goal: Personalized ranking of stock funds
• Characteristics: less risky than stocks, low volume, sparse data
Business plan: a formal statement of business goals, reasons they are attainable, and plans for reaching them
• Goal: Personalized recommendation of questions
• Characteristics: Low risk, explanation is not desired
10. 12
Domain-based review / Online banking and multi-domain solutions
• Unified user interface to access items from various domains
• Goal: Personalized user interface, asset allocation
• Applications:
› FSAdvisor: Personalized sales support for representatives in financial services
› Context-aware credit card-based mobile recommender
› Personal Choice Point: Personal financial planning tool
› Investment decision support system
Heterogeneity Churn Rate Int./vol. Preference Risk Scrutability
high low explicit unstable high high
11. 15
Domain-based review / Conclusion
• Identified a number of domains (additional ones?)
• Financial domains are risky, specific
• Found a number of real-life applications for microfinancing, insurance and real estate
• Huge literature for non-personalized stock and portfolio recommendation, smaller one for personalized
applications
• Promising concepts for venture finance, stock fund and business plans
• Several concepts for multi-domain service
• Significant willingness to introduce decision support systems to banking environment
13. 17
Method-based review / Collaborative Filtering
• Principle: Interaction-based consumption patterns, user or item based similarity
• Pros:
› Universal method, does not require complex data
› Consumption patterns, unexplainable preferences
• Cons:
› Difficult to explain
› Inefficient for complex problems
› Cannot handle cold start
• Financial domains: loan insurance riders, real estate, venture capital, stock market
14. 18
Method-based review / Content-based Filtering
• Principle: Metadata matching/similarity-based recommendations
• Pros:
› Simple, but can be extended for multi-criteria-based recommendation
› Easy to explain
› Handles item cold start problem
• Cons:
› Not typical in financial domains (fuzzy)
› Difficult to model multi-criteria
› Less efficient
• Financial domains: real estate
15. 19
Method-based review / Knowledge-based recommendation
• Principle: Knowledge acquisition and representation, criteria-based recommendations
• Pros:
› Acquires specificity of domains
› Able to handle complex problems
› Easy to explain
› Handles user cold start problem
• Cons:
› Knowledge acquisition process
› User interface
• Financial domains: loan, multi-domain solution, business plans
16. 20
Method-based review / Case-based recommendation
• Principle: Case-based reasoning (CBR), solving recommendation problem based on on past cases
• Pros:
› Efficient for multi-criteria searching problems
› Avoids knowledge acquisition
› Handles user cold start
• Cons:
› Sufficient amount of criteria and cases
› Finding similar cases is not trivial
• Financial domains: insurance, real estate, P2P lending, asset allocation
17. 21
Method-based review / Hybrid and complementary methods
• Hybrid methods: Combination of single methods
› Collaborative- and content-based filtering (insurance, stock, stock fund)
› Context-awareness hybrid solutions (multi-domain)
› Hybrid association rule mining (insurance, stock)
› System components: social, economic and semantical information extraction agents (stock)
• Complementary methods
› Fuzzy: Preference modeling (real estate, stock, portfolio management)
› Artificial neural networks: extracting information from news, forecasting stock price, semantic analysis
› Classification: Support Vector Machines (stock market – forecasting, buy/sell signals)
18. 22
Method-based review / Conclusion
• Collaborative-filtering is mainly used for stock market
• Content-based filtering is not typical, only for hybrid methods
• Knowledge-based-methods are suitable, but we found a few applications only
• Case-based recommendation is efficient, but it requires decent amount of data
• Context-awareness is less studied yet
• Complementary methods are mainly used for stock market