Special Challenges With Large Data Mining Projects CAS PREDICTIVE MODELING SEMINAR Beth Fitzgerald ISO October 2006
Agenda Project Overview Prior to Modeling Modeling Business Issues
Development of a Model - Project Overview Data Statistical Tools Computer Capacity Team Skills – Data management – Analytical/statistical – Technology – Business Knowledge
Prior to Modeling Formulate the Problem Evaluate Possible Data Sources Prepare the Data Develop Understanding of Modeling Procedures and Diagnostics Explore the Data with Simple Modeling Techniques
What percent of a model building project is the data preparation and data management? 25% 50% 75% 85%
Formulate the Problem What problem are you trying to solve? What results do you expect to see? How will you know if the results are reasonable?
Prepare the Data Do quality checks in level of detail needed for project Understand how to prepare individual variables for use in models Need to be practical about number of classification categories models can handle Need to decide on truncation and bucketing of variables that are continuous Create new variables
Develop Understanding of Modeling Procedures and Diagnostics Basic modeling training – GLM, Data Mining What software is available? What software/models work for my data investigation, modeling problem, etc. What computer capacity do I need? Learn how to use software Learn how to interpret the diagnostics
Development of a Model Analyze historical policy and loss data – Policy level detail – Location level detail Link policy and loss data with external and/or internal data: – Specific business risk data – operational, financial – Specific location data – demographic, weather – Other data – building, vehicle, agency Need link between policy detail and other data
Explore the Data with Simple Modeling Techniques Start with sample of data Try different classical analysis on sample such as: – regression – linear models – correlation matrices Make use of graphical options to explore data
Data Management Issues Matching additional internal policy information to premium/loss data – Different points in time – Tracking & balancing audited exposures Different summarization keys – handling of mid-term endorsements Address scrubbing Matching to external data for correct point in time Significance of missing values within variable
Modeling Activities Selection of Predictors – variable elimination, variable transformation Start with classical models prior to evaluating more complex models Methodology Understanding and Evaluation Evaluation of Model Performance
Data Mining Techniques Balance good fit with explanatory power Generalized Linear Models Classification Trees Regression Trees Multivariate Adaptive Regression Splines Neural Networks
Data Mining Process Business Knowledge Data Linking Data Cleansing Analyze Variables Determine Predictive Variables Evaluation Data Gathering Data Mining
Model Performance Lift Curve Analysis – Score all risks in sample – Rank risks by score from Bad to Good – Compare loss ratio of risks in each decile to loss ratio for all risks
Sample Lift Curve Analysis
Business Issues Model uses information from a third- party vendor Model needs to be accessible electronically Technology Issues Implementation Decisions
Technology Issues Develop/Modify Systems Integrate into underwriting/rating workflow – Decision process – Agency system Decide on technology – Web-based interface – API, FTP, MQ, TCP/IP, HTTPS webservices
Implementation of Model Solution focus/usage: Suitability of risk for underwriting decision Source for additional pricing factors Consistency in underwriting/pricing decisions Compliance with regulations based on implementation decision Consider model alone or model with other information available from application
Implementation of Model Workflows: Underwriting – New Business – Renewal business Rating – Pricing – Coverage Adjustment
Business Implementation of Model Strategic Plan - need management involvement Prepare Announcement/Training Material for Internal & External Customers Coordinate Implementation Monitor Feedback/Adjust Implementation
Future Plans Determine Process for Updates to Model – Use of Updated Data – Use of New Data Variables – Use of New Techniques