The Business of Predictive Modeling December 17, 2013 Christine Hofbeck, FSA, MAAA Centroid Analytics, LLC AGENDA PART I -- INTRODUCTION PART II – MODELING 101 (Basic Steps) PART III – “GOLDEN QUESTION” PART IV – OPERATIONAL CONSIDERATIONS 2 INTRODUCTION Predictive modelling [sic] is the process by which a model is created or chosen to try to best predict the probability of an outcome. -- Wikipedia VISUALIZE In practice: Identify patterns/ segment risks 3 Customers Most Profitable Lines Develop business rules Improved decision OTHER making you are only limited by your creativity Potential Applications OPTIMIZE VISUALIZE Operational Efficiency Distribution Channels Claims Management Pricing / Reserves Customers Most Profitable Lines or Products Target Marketing DATA MINIMIZE Risk Fraud 4 OTHER you are only limited by your creativity Potential Applications (Life) 1. Triage UW decisions; implement STP for (more) applicants 2. Decrease purchase of traditional UW requirements by determining when they may not be necessary 3. Identify & target customers more likely to buy 4. Identify customers more likely to lapse – intervene if profitable, allow unhealthies to lapse 5. Inforce book management 6. Identify most desirable agents 7. Smart customer handling 5 Predictors A predictive model is made up of a number of predictors (“independent variables”), which are data elements likely to influence future behavior or results (“dependent variable”). DON’T USE ONE VARIABLE the mean predicts the future but doesn’t tell us why… (“underfit”) 6 DON’T USE ALL VARIABLES exactly replicates the past… cannot predict the future (“overfit”) SEEK PARSIMONY BASIC STEPS (Modeling 101) Define & Scope Review & Refine Data Prep Model Build Implementation 7 Model Validation Define & Scope Exactly what are we trying to predict? For whom/what are we trying to predict this? (“unit of exposure”) Do we have the systems capacity to implement? 8 How will the results be used? What is the budget? Consider IT, staff, data purchase, training, etc. How long do we have to build? To implement? Insource or outsource? Data Prep * sometimes the most time intensive step of modeling INTERNAL DATA o o o o # years accuracy ability to access primary key EXTERNAL DATA o o o o match rate cost – to model cost – to use frequency of update MODELING DATASET 9 Consider both expected & unexpected relationships – creativity in data exploration can be the key to your competitive edge! Data Prep (cont’d) 1. COMBINE various data sources 2. CONVERT to desired exposure unit or format 3. CORRECT inaccurate data 4. INSPECT to remove variables: - Too many blank values that cannot be imputed - All/most values the same Data cannot be relied upon Data will not be captured going forward Legal advice not to use 5. BUCKET (“bin”) values 10 Model Build (cont’d) UNIVARIATE ANALYSIS – test each variable one by one to see which ones may be predictive. MULTIVARIATE ANALYSIS – examine multiple variables in different groups to obtain the best, USABLE results – remember parsimony! INTERACTIONS – which variables can be combined into a “mega variable” to improve results (i.e., does 1+1 = 1.5? does 1+1 = 3?) 11 Complicate the model (add variables, interactions) and simplify the model (remove variables, bin) to find the preferred combination. Model Build (cont’d) Various tests can be used to determine variable inclusion: 12 STATISTICAL CONSISTENCY JUDGMENT P-values Cramer’s V Confidence intervals Type III tests Of patterns - Apply business knowledge to assess whether suggested relationships make sense Over time Over random parts of a dataset Model Validation ACTUAL vs. EXPECTED-- how close did we get? Generally, a subset of the data is withheld during the modeling process for validation: OUT OF TIME OUT OF SAMPLE withhold most recent data withhold randomly generated % of records Model validation graphs are useful for communicating model performance to non-technical audiences. 13 Model Validation – Sample Chart 0.9 0.8 Outcome 0.7 0.6 Actual 0.5 Expected 0.4 0.3 0.2 1 2 3 4 5 Decile 14 6 7 8 9 10 Implementation BUSINESS RULES What decisions will be made based on the prediction? May vary by location, business, rate group, etc. SYSTEM BUILD Scoring engine (collects data & calculates predictions) Decision tool (executes business rules) User interface TRAINING Anyone who will interact with the model must understand what it does and why 15 Review & Refine REPORTING How close did we get to the goal? How far did we exceed it? Multiple reporting packages required for varied audiences, for example: Executives – highlights in aggregate by zone, business unit, product Actuaries – detailed results by variable, state, rate group Marketing – by broker/agent, location Underwriting – by underwriter as a performance measure Frequency of update – weekly, monthly, quarterly, yearly? Method of calculation – automated? ad hoc? Review & Refine (cont’d) MODEL UPDATES WHY? As target customer is attained, characteristics of inforce book will change Business goals/strategies may change New data may become available Tolerance for certain characteristics may change HOW? Update current variable relativities (“recalibrate”) Start over - search for more predictive variables (“recast”) HOW OFTEN? 17 Advantages of Modeling Over Traditional Approaches 1. Many additional and often unconventional variables may be examined 2. Modeling a particular variable controls for the effects of other included variables – we don’t risk double counting or attributing effects to the wrong variables 3. Traditional approaches segment data into smaller categories which impact credibility 4. Interactions are introduced 18 The above advantages can lead to improved accuracy, enhanced business and strategic benefits, more reliable assumptions, improved risk mitigation, etc. THE “GOLDEN QUESTION” Through brainstorming, feedback loops, and data review, determine what single characteristic (“golden question”) will define your target 19 OPERATIONAL CONSIDERATIONS 1. Executive & cross-functional support 2. Time/cost versus depth of investigation 3. Strategic modeling process 4. Cross-functional involvement throughout build 5. Thorough training 20 Executive & Cross-Functional Support If target users don’t support the model, they will resist using it. Gaining complete support can be difficult: 1. Resistance to change 2. Concern that model results will highlight current deficiencies 3. Lack of understanding of predictive models 21 Support (cont’d) My position will be eliminated if a model is now used to select risks. My expertise must not be important to the company. The model will suggest that my current method is incorrect, which will reflect poorly on my performance/reputation. I will have to take on additional work associated with new processes. My workflow will double (triple). 22 I already have an established plan. I know who our target customer is. We’ve always done it this way, and it’s worked. I don’t see a reason to change anything. I don’t know how to explain this to a broker/agent so I don’t want to use it. I found one outlier so the model must be wrong. Time/Cost vs. Depth of Investigation The process of building and implementing a model can typically be quite lengthy – longer than most expect • • • • • • Simpler Study (3-12 months) Results more conservative Perhaps internal data only Generous binning Limited interactions May be appropriate if goal is a general sense of direction 23 OR • • • • More thorough investigation Additional time Additional development cost Possible greater payoff through enhanced segmentation and data exploration Remember that a simple model does not necessarily indicate a simple study! Strategic Modeling Process TARGET PREDICTION/USE o Ensure target is appropriate for the intended use o While many ideas are interesting, you may wish to focus on those which are actionable STATISTICAL SIGNIFICANCE vs. ULTIMATE IMPACT o The most statistically significant model may not be the most impactful o Consider ease of implementation, repeatability, updates o Identify when “less is more”! FLEXIBILITY o Allow for unexpected insights which could lead to unanticipated changes in business strategy or process o Sometimes the insights gained from the journey will prove more important than the planned goal Cross-Functional Involvement Data, product & IT experts, legal advisors, and model users must remain engaged throughout the model build Insight from functions Insight to functions Eases training and implementation Keep modelers apprised of changes in strategy 25 Legal considerations around certain variables Thorough Training The model isn’t done when it’s done. Who will provide the training? Who is most appropriate to provide training? Modeling team General training team Functional experts Consulting team* Other No clear answer – but this must be thoughtfully considered and appropriately executed to reap the full benefits of the model which was built 26 *Consider what information may be shared (non-proprietary) Discussion/Q&A Remember… Modeling is a complete business strategy NOT just a mathematical process So how will YOU use predictive modeling to improve your business? 27 Christine Hofbeck, FSA, MAAA Centroid Analytics, LLC christine.hofbeck@centroidanalytics.com 908.884-4103 (c) 908.574-5351 (w) www.centroidanalytics.com