Modelling the 3600 Customer View Neill Crossley Consulting Director, Analytic Solutions FICO 16 November, 2010 Confidential. This presentation is provided for the recipient only and cannot be reproduced or shared without Fair Isaac Corporation's express consent. 1 © 2010 Fair Isaac Corporation. Agenda » Introduction to Predictive Analytics » Trends in Data » Trends in Predictive Models » Benefits of Frictionless Model Deployment » Summary 2 © 2010 Fair Isaac Corporation. Confidential. © 2010 Fair Isaac Corporation. Confidential. Introduction to Predictive Analytics 3 © 2010 Fair Isaac Corporation. Confidential. What is Predictive Analytics? “Predictive Analytics encompasses a variety of techniques from statistics, data mining and game theory that analyze current and historical facts to make predictions about future events.” “One of the most well-known applications is Credit Scoring, which is used throughout financial services. Scoring models process a customer’s credit history, loan application, etc., in order to …rank-order individuals by their likelihood of making future credit payments on time.” 4 © 2010 Fair Isaac Corporation. Confidential. Predictive Analytics Give You Better Predictions, Better Decisions Descriptive Models Profiling & Segmentation Predictive Models Predictive “Scores” Decision Models Decision Modelling & Optimisation Decision Strategies Economic Forecast p(able to pay) Bureau Data (FICO Score, Bureau debt) p(foreclose) Monthly Payment Income, household size, etc. Current loan amount, interest rate and term PV(Loss | foreclosure) Interest Reduction; Principal Reduction; Term Externsion p(walk away) NPV PV(Revenue | good) Client predictive models and other key data Future Value of Home Time to foreclosure Current Value of Home Region Establishes broad segments based on customer profile data 7 © 2010 Fair Isaac Corporation. Confidential. Rank orders prospects / customers on a single dimension Combines Predictive Models with business rules to create Micro Segments Complete Decision Framework Assigns the best action given specific business constraints Achieving a 3600 Customer View ASSETS CHARACTER INCOME Current Account Overdraft ABILITY Savings Credit Card Accounts Accounts BORROWING 3600 Basel II PREFERENCES Customer View Retirement Accounts Mortgage Accts Installment Loan Non Financial Transactions INTERESTS LIFESTAGE SECURITY 8 © 2010 Fair Isaac Corporation. Confidential. COMMITMENTS Analytics Impact Business Results … at All Points of the Life Cycle Predictions Decisions Acquisition 9 Whom to target What product to offer Approve/decline Line management Collection priority Credit amount Re-pricing/renewal Collection action Price Transaction authorizations Channel placement Agency placement Channel Features Timing Up-sell Capital Response Revenue Risk Capacity © 2010 Fair Isaac Corporation. Confidential. Collections & Recovery Customer Management Origination Cross-sell Capital Risk (Loss/BK) Risk (Loss/BK) Bankruptcy Revenue Revenue Charge-off Capacity Capacity Roll Pre-payment Pre-payment Response to an up-sell Response to a cross-sell Expected collection amount Fraud Fraud Trends in Data 10 © 2010 Fair Isaac Corporation. Confidential. Predictive Analytics Data – 10 + years ago Acquisition Origination Customer Management Demographic x X X Negative Credit Bureau x X X x x X x Account Behaviour Limited Access to Clean Accurate Data Predictive models were built using <100 variables 11 © 2010 Fair Isaac Corporation. Confidential. Collections & Recoveries Predictive Analytics Data - Now Acquisition Demographic Negative Credit Bureau x Origination Customer Management X X X X x X x X X X X X x x x X x x X Account Behaviour Customer Behaviour Positive Credit Bureau Financial Transaction Fraud Deposit Data Offer & Take up Collections Non Financial Transaction Geo-demographic X x X X X X X X X X x X x x x Collections & Recoveries x X x x Model Developers often have hundreds and thousands of potential variables 12 © 2010 Fair Isaac Corporation. Confidential. More Data Means Typical Data Challenges Become a Bigger Issue 13 © 2010 Fair Isaac Corporation. Confidential. Clean Data Deploy Models Create Characteristics Develop Models Problem – Amount & Correlation of Data » Correlation has always been an issue in model development IV = 3 IV =2 IV=1 What about the impact of correlation? Which two Characteristics are best? MC=1 MC=2 » Increased Problem: more data = more potential for correlation » Can impact some statistical techniques » Leading to unusual weight patterns » Which of the Fully or Heavily Correlated Variables should you use? » In what way are they correlated? 14 © 2010 Fair Isaac Corporation. Confidential. Problem – Amount & Correlation of Data » FICO Model Builder Score Engineering technology » Through use of Score Weight Constraints & Optimization algorithms » Patterns, Grouping, In-weighting, Cross-Restrictions, No-Inform » Entry-Exit based on Marginal Contribution to Divergence » Interactive what-if analysis » Variables Tiers – stepwise and diversity variable selection » Looking at interactive functionality to further improve variable reduction 15 © 2010 Fair Isaac Corporation. Confidential. FICO Model Builder uses unique technology building blocks Fitting Objectives Score Formula (Segmented) Generalized Additive Model Capture nonlinear relations and interactions accurately whilst being interpretable by business users and regulators 16 © 2010 Fair Isaac Corporation. Confidential. Divergence, Range Divergence, Bernoulli Likelihood, Multiple Goal, Tune score weights to optimize and trade-off business objectives Score Weights Constraints Patterns, Grouping, In-weighting, Cross-Restrictions, No-Inform Impute domain knowledge to obtain palatable and robust models Optimization Algorithms Various Solvers, Goal Programming, Automatic Variable Selection Quickly and reliably solve for optimal model. The FICO approach helps overcome the multi co-linearity issues seen with standard regressions. Trends in Predictive Model Types 17 © 2010 Fair Isaac Corporation. Confidential. Trends in Predictive Models Types Attrition / Retention Models » “Customer retention has returned as a core value to financial services institutions. Analytics can play a major role.” - Tower Group Customer Level Models » Use all available customer information across a range of different products, to measure the likelihood of a specified future event, in a given time horizon. Action Effect Models » Action Effect models predict the outcome based on impact of the different actions that could be undertaken e.g. how customers react to different collections actions. Capacity / Affordability Models » Anticipatory risk measures that rank orders customers according to their ability to manage new or increased debt safely on top of their existing debt commitments 18 © 2010 Fair Isaac Corporation. Confidential. Use of Transaction Scores Positives » There’s lots of Transaction data » Captures detail about a customer’s habits, lifestyle etc » Can help you make more accurate and timely decisions Negatives » There’s lots of Transaction Data » Difficult to store over time » Difficult to operationalize & process in decision management » Not always clean or detailed enough on all accounts FICO has been modeling and operationalizing transaction-based credit risk scores since 1996 » FICO® Transaction Scores™ (FTS) produces a new score for every transaction so changes in risk profile are picked up quickly » Based on FICO patented transaction profiling technology used in FICO™ Falcon » Results in average 6% improvement in predictive power over behaviour score 19 © 2010 Fair Isaac Corporation. Confidential. FICO Transaction Scores UK Consortium Model Performance (KS) % of Total Accounts % of Total Bads FITS + Behavior Score KS Behavior Score KS % Lift in K-S Current Very Mature w/ Cash (>6 years time on books) 9.2% 8.8% 62.15 54.68 13.7% Current Mature w/ Cash (7 11.4% 20.1% 48.35 42.51 13.7% Current Young w/ Cash (<7 2.2% 6.1% 49.28 44.06 11.9% Current Mature w/o Cash w/ Frequent Transactions 52.0% 15.6% 67.47 61.88 9.0% Current Mature w/o Cash w/o Frequent Transactions 18.3% 8.4% 60.92 57.01 6.9% Current Young w/o Cash 4.5% 2.0% 48.59 38.19 27.2% Delinquent 1 Cycle 2.0% 22.3% 48.75 45.59 6.9% Delinquent 2 Cycle 0.4% 16.7% 39.88 38.04 4.9% Overall 100% 100% 70.67 66.70 6.0% Segment mo – 6 years time on books) mo on books) 20 © 2010 Fair Isaac Corporation. Confidential. Based on FITS UK Consortium Models v.1.0 – Out-of time validation results (Sep ‘08, Mar ’09, Sep ‘09) FICO Transaction Scores – Timing Impact in identifying Bad accounts A larger amount of bad accounts were triggered first by the transaction score, than by the cycle score Transaction scores triggered first Cycle scores triggered first Percentage of 3+ Cycle Accounts Triggered at the same time 21 100% 80% 60% 72% 65% 40% 20% 5% 6% 60% 7% 53% 6% 29% 34% 41% 23% 5% 10% 15% 20% 0% © 2010 Fair Isaac Corporation. Confidential. Score Cutoffs - Predicted Bad Rate Model When Things Will Happen – Time to Event Issue » Traditional models focus on whether something is likely to occur, not when » Significant information from existing data being overlooked » May significantly impact key business decisions: » When to increase a limit » When to offer another product » Whether to accept a loan application Example: » A low risk Instalment Loan applicant with no other relationship with the Bank » Is predicted to repay in full within 4 months and thus may actually be value destroying FICO has been pioneering Time to Event Models over the last 5-6 years » Have been successfully utilised in innovative work with a number of US Retailers – large increases in offer take up rates » Showing potential applications in Financial Services: » » » » » 22 Predict time to default? Predict time to closure? Predict mortgage prepayment? Predict duration of a call center call? Predict time to up-sell a product enhancement? © 2010 Fair Isaac Corporation. Confidential. Model Economic Conditions - Economic Impact Service Issue “We have been driving a car by looking out the rear view mirror” » Models have been built ignoring the impact of past economic conditions » And how these may change - “The future is never exactly like the past” Requirements going forward » “Be anticipatory, not just reactive” “Incorporate a forward looking view of our business” » Regain control of your portfolio » Increase confidence in your risk management tools, even in these times of change FICO Economic Impact Service » Helps translate how current and projected market conditions will impact the expected risk levels associated with given score bands » Clients can more scientifically adjust strategies to reflect the upcoming changes in risk » Thus can help: » Limit losses; Grow portfolio responsibly; Better prepare for the future; Meet regulatory compliance » Broadly applicable to and easily integrated with a variety of score types and uses 23 © 2010 Fair Isaac Corporation. Confidential. Model Impact of Decisions – Decision Models Issue » Decisions made by Financial Services company can significantly impact the outcome » Lend too much, customer may not be able to afford, increasing likelihood of default » Not lend enough, customer may go to a competitor Requirements going forward » Understand the impact of different decisions on customers » Understand all impacts of different decisions on the components of Profit » Then identify the Sets of Decisions that bring the most profit but also do not break key business goals and constraints e.g. Annual Portfolio Losses can not exceed €50m. » Measure against expectations FICO has been a pioneer in Decision Modeling & Optimization over the last 10 years » FICO’s approach provides an Open Framework & Methodology for modelling decisions. »Builds on existing Scores & Systems »Focused »Refined »Proven 24 on Business Goals & Constraints on over 100+ projects Business Benefits © 2010 Fair Isaac Corporation. Confidential. Post-Crisis Predictive Analytics Summary Data » Utilise all relevant internal and external data » But be practical – cost v benefit » Look out for tools and techniques to help: » Identify the most relevant data » Operationalize data » Focus on breadth of Data » Experimental Design / Learning Strategies Models » Model the Decision to optimize performance » Factor in Macro Economic forecasts » Simulate Strategies to understand potential future impacts & set expectations » Balance Automation with Expertise » Justify complexity and increase transparency » Look to understand the 3600 view of the customer, but from their perspective 25 © 2010 Fair Isaac Corporation. Confidential. Post-Crisis Predictive Analytics Summary Some Phrases to consider: “Don’t’ tell me what you think, but what you know” “If you can’t measure it, you can’t manage it” “ All Models are wrong, some are useful” “Show me the data, explain why I should trust the model and tell me where its weaknesses are that I need to be aware of”. “No one doubts that more data and more relevant data leads to better models. The winners, however, won’t just use that data to build better models – they will use it to ask better questions.” Dr Andrew Jennings, Chief Research Officer, FICO 26 © 2010 Fair Isaac Corporation. Confidential. Benefits of Frictionless Model Deployment 27 © 2010 Fair Isaac Corporation. Confidential. Deploying Models Is Not Always Simple » Traditional technique » Document model » Ask IT to recode » Lengthy testing Predictive Model Specs IT Software Development SQL in Database » How many £ lost per day? Time To Deploy Predictive Models Self-Reported Measures, Global Financial Companies FICO Survey 2008 % responders 40 30 20 10 0 1-4 weeks 28 © 2010 Fair Isaac Corporation. Confidential. 1-3 months Compiled Executable 3-9 months 9 or more months Frictionless Model Deployment Has Developed Over the Last Few Years FICO Model Builder » Imported models are executed as Java components in the decisioning service. Rule developers and business users cannot see nor modify them Rule Service Decision Management Repository Rule Service » Imported models are available to rule developers and authorized business users can see and modify them » PMML integration 29 © 2010 Fair Isaac Corporation. Confidential. Java .NET Code Gen COBOL FICO Model Builder Supports Parallel Development & Deployment Approach FICO Model Builder Production Data Translate Metadata Execute Variables Execute Model Test Model Deployment Environment Direct Deployment Other Data Sources Data Load & Transform Other Modelers 30 © 2010 Fair Isaac Corporation. Confidential. Create Variables Share Build Model Modeling Repository Validate Model Rebuild model Monitor Validate Which can be further extended to create the FICO Analytic Platform Lowers your lost opportunity and model management costs Features Centralised Model Repository Across Geographies, Portfolios, Decision Areas Faster model deployment Models editable in production Central Characteristic Library asset Model lifecycle management Model monitoring and reporting Simple reuse, redeployment and refresh Model and Decision simulation 31 © 2010 Fair Isaac Corporation. Confidential. What is an Analytic Platform? An integrated solution that can manage the whole predictive model lifecycle. Data Preparation Development Prioritisation Model Realign, Reweight & Refresh Manage Development Assets MODEL LIFECYCLE Lifecycle Management Model Monitoring Manage Use in Production 32 © 2010 Fair Isaac Corporation. Confidential. Development Testing, Validation & Simulation Deployment THANK YOU Confidential. This presentation is provided for the recipient only and cannot be reproduced or shared without Fair Isaac Corporation's express consent. 33 © 2010 Fair Isaac Corporation.