2007 Predictive Modeling Seminar Handouts

These handouts are available in their original Power Point Presentation format. If you do not have Power Point on your computer, you will need to download the free Power Point Viewer, which is a 2.8 MB file.

Some of the handouts are Portable Document Format (PDF) files. The files were created using Adobe Acrobat. To view the documents, you need the Acrobat Reader. For more information, read the Adobe Frequently Asked Questions. To download the free Acrobat Reader, click the icon below.

Audio recorded sessions are also available online for select sessions below.

If a session handout is not online, unfortunately it was not submitted by the speaker and is not available.

General Session/Keynote Speaker

John F. Elder IV
Chief Scientist, Elder Research, Inc.

Download Audio

Download Handout

Dr. John Elder heads a data mining consulting team with offices in Charlottesville, VA, and Washington, DC. Founded in 1995, Elder Research, Inc., focuses on scientific and commercial applications of pattern discovery and optimization, including stock selection, image recognition, medical text mining, biometric identification, drug efficacy, credit scoring, cross-selling, investment timing, and fraud detection.

Dr. Elder obtained a BS and MEE in electrical engineering from Rice University, and a Ph.D. in systems engineering from the University of Virginia, where he’s recently been an adjunct professor, teaching optimization. Prior to a decade leading ERI, he spent five years in aerospace defense consulting, four years heading research at an investment management firm, and two years in Rice’s computational & applied mathematics department.

Dr. Elder is an author of innovative mining tools and is active on statistics, engineering, and finance boards, as well as with conferences in these fields. A frequent keynote conference speaker, he will serve as general chair of the 2009 Knowledge Discovery and Data Mining conference in Paris. Dr. Elder’s courses on data analysis techniques—taught at dozens of universities, companies, and government labs—are noted for their clarity and effectiveness. Dr. Elder was honored to serve five years on a panel appointed by the president to guide technology for the National Security Agency.

Concurrent Sessions

ADVANCED MODELING TECHNIQUES

Download Handout

This session will explore a variety of useful techniques not covered in other sessions. A common theme will be reflecting nonlinear patterns and variable interactions in ones model. The session will feature a discussion of tree-based modeling, as exemplified by the “Classification And Regression Tree” (CART) algorithm. While CART can be used as a predictive modeling tool, it can also be used for variable selection, variable binning, exploratory data analysis, data visualization, and model analysis. Additional topics to be covered will include Generalized Additive Models (GAMs), Neural Networks, and MARS (Multiple Additive Regression Splines). Emphasis will be placed more on concepts and working intuition than mathematical formalism. The techniques’ relative advantages and disadvantages will be discussed, and relationships of the various techniques with regression and Generalized Linear Models will be pointed out.

Panelists:

COMMERCIAL AUTO PREDICTIVE MODELING: THE SAME AND COMPLETELY DIFFERENT

Download Audio (.mp3)

The commercial automobile insurance product line offers a unique predictive modeling opportunity. On one hand, many of the factors and data sources developed for personal auto can be used in commercial auto. Credit scores, vehicle identification numbers (VINs), motor vehicle records (MVRs), and territory refinement all make the transition to commercial auto quite well. On the other hand, the unique characteristics of commercial auto risks offer some new opportunities. Factors such as industry classification, trailer type, and truck-to-car ratio all contribute to the dynamic world of commercial auto predictive modeling.

Panelists:

COMMERCIAL LINES PREDICTIVE MODELING – BOP

Download Handout 1
Download Handout 2

In an attempt to replicate the successful applications in the personal lines industry, the commercial lines industry is speeding up its adoption of predictive modeling. This session will review the current predictive modeling development for small commercial risks. The session will also discuss the underwriting and pricing challenges for small commercial risks and how predictive modeling, such as scoring models, can address these needs. The efforts and challenges involved in building predictive models will be described, including data issues and analysis of models.

Panelists:

COMMERCIAL LINES PREDICTIVE MODELING FOR WORKERS COMPENSATION

Download Handout

Predictive modeling is proving to be as successful in commercial lines as it has been in personal lines. Many of the challenges that face commercial lines modelers, however, are different and must be addressed properly in order for the results to be reliable.

This session will discuss the major challenges in workers compensation modeling, from data to deployment, and will offer approaches to address these challenges.

Panelists:

CREDIT SCORING UPDATE

Download Handout 1
Download Handout 2

The use of credit in insurance continues to be a topic that is hotly debated by insurers, consumers, regulators, and other federal and state government officials. A major challenge for the industry is that there is not an understanding of why there is an association between scores and loss propensity. The panel will discuss a paper by Brokett and Golden that deals with this issue, “Biological and Psychobehavioral Correlates of Credit Scores and Automobile Insurance Losses: Toward an Explication of Why Credit Scoring Works.”

The second part of the session will cover the much anticipated and recently released Federal Trade Commission study of auto insurance scoring and its effect on different segments of the population. The study is based upon insurance data provided via the cooperation of three major trade associations, and data from the Social Security Administration and third-party data providers. A brief summary of reactions to the study by interested parties will also be covered.

Moderator:

Panelists:

DATA QUALITY—RAISING YOUR ACTUARIAL IQ (IMPROVING INFORMATION QUALITY)

Download Handout 1
Download Handout 2
Download Handout 3

Predictive modeling, Sarbanes-Oxley, and other recent developments have renewed the focus on the quality of information. This session starts with some striking examples of the cost of poor information quality, defines the concept of data quality, and then gives tips and examples on how to pursue it, including how actuaries can be proactive in improving data quality. The emphasis will be on:

Techniques that should be easy for most actuaries and analysts to apply right away.
Aspects of data quality that actuaries are best able to fulfill.

This session is drawn from the work of the CAS Data Management Educational Materials Working Party.

Predictive modeling often coincides with the first time insurance data is looked at very closely. This is often overwhelming, as unknown data problems are sometimes uncovered during the process. This presentation will talk about common insurance data problems, how they can be identified before the modeling begins, and how they could possibly be solved.

Panelists:

DIMENSION REDUCTION

Download Handout

A simple predictive model should be favored for implementation over a complex model. Predictive modeling creates a framework where the actuary can comprehensively study a large amount of data from numerous sources, and variables with many distinct values. As a profession we would like to understand all of the underlying drivers of losses. However, it is not possible or practical to include all possible variables, or all possible values of a variable, in the final rating model. This session will discuss different approaches and reasons to reduce the number of variables or reduce the number of values for modeling.

Panelists:

Estimating Personal Auto Loss Costs that Vary by Address/Household Averaging

Download Audio (.mp3)

Download Handout

Postal zip codes form the basic unit for many territorial ratemaking methodologies that are in use today. Driving conditions such as chronic traffic congestion, population density, weather, and the physical environment are not constrained by zip code boundaries. The first part of this session describes how to estimate personal auto loss costs as a function of over 1,200 variables that describe local driving conditions.

The method proceeds by first applying a number of variable reduction techniques, such as principal components analysis and structural equation models, to significantly reduce the number of variables. Then it fits separate frequency and severity models based on this reduced number of variables to produce loss cost estimates. The session will then describe how to analyze a holdout sample and measure the effectiveness of this methodology.

Historically, operators in personal lines auto have been assigned to vehicles using outdated underwriting standards. The challenge has always been in reflecting other drivers in the household and there potential for using the insured vehicle. Over the past several years, insurers have been using predictive modeling techniques to better incorporate the information about the extra operators on to the vehicle. This second part of the session will discuss several strategies associated with this challenge within a predictive modeling framework.

Panelists:

FREE AND CHEAP DATA SOURCES

Download Handout 1
Download Handout 2
Download Handout 3

It is widely believed that the usefulness of predictive models can be increased by incorporating external sources of data, along with company-specific data, into the project database. This session will feature a number of short presentations by users of external data. The data featured in this session is either free or costs no more than $200 to acquire. The presenters will include information about geographic, demographic, and economic and survey data. Attendees at this session will learn how to find the data and will be provided with some examples of applications of the data.

Moderator:

Panelists:

FREQUENCY AND SEVERITY MODELING

Download Handout

Most predictive modelers and software packages use the Poisson and Gamma distributions for frequency and severity modeling. During this session, we will consider alternative distributions to the Poisson and Gamma and describe how these choices can affect parameter estimates. We will illustrate other useful conditional distributions for severity modeling, including one distribution that is essentially absent from the actuarial literature.

Alternative modeling strategies, such as modeling the pure premium directly, will be covered. We will discuss the impact of data collection and coverage (clustering of loss values, limit, and deductible) on the model, and how to adapt the model for these situations.

Panelists:

GLM I: INTRODUCTION TO GENERALIZED LINEAR MODELS

Download Handout
Download Handout (.xls)
Download Handout

Do terms such as “link function,” “exponential family,” and “deviance” intimidate you? If so, this session will help demystify generalized linear models and will provide a basic introduction to linear and GLMs. It is targeted at those who have modest experience with statistics or modeling.

The session will start with a brief review of traditional linear models, particularly regression, that have been taught and widely applied for decades. Session leaders will explain how GLMs naturally arise as some of the restrictive assumptions of linear regression are relaxed. GLMs can model a wide range of phenomena, including frequencies, severities, and loss ratios, as well as the probability that a customer will renew a policy, to name just a few. The session will emphasize intuition and insight rather than mathematical calculations, which are handled by software these days, anyway!

Panelists:

GLM II

GLM I provided the case for using GLMs and some basic GLM theory. GLM II will be a practical session outlining basic modeling strategy. The discussion will cover topics such as overall modeling strategy, selecting an appropriate error structure and link function, simplifying the GLM (i.e., excluding variables, grouping levels, and fitting curves), complicating the GLM (i.e., adding interactions), and validating the final model. The session will discuss diagnostics that help test the selections made.

Panelist:

GLM III

GLM III will cover additional refinements, such as investigating the appropriateness of a multiplicative model structure, how to combine GLMs across multiple claim types, and the use of the offset term to constrain models. The session will also consider techniques for modeling large claims, and practical model validation approaches. It will also consider specific issues that arise when modeling price demand elasticity with GLMs, which is of particular importance when undertaking price optimization analyses.

Panelist:

GLM Offset/Geneneralized Iteration Algorithms

Download Handout 1
Download Handout 2

The first part of the session will focus attention on the offset feature of GLMs. The offset feature gives the modeler control over the weight selected terms have in a model. The panelists will discuss various ways in which the offset feature can be creatively used to meet actuarial challenges in personal and commercial lines modeling projects.

The second part of this session will present a very flexible and comprehensive iteration algorithm called generalized iteration algorithms (GIA) to build up GLM models. The algorithm is a generalization of the minimum bias approach. It will be demonstrated that GIA can be used to solve all commonly used GLM models.

There are three main advantages for GIA. First, it can solve not only the commonly used GLM models, but also a broad range of GLM models. Therefore, it gives actuaries more options to fit their data. Second, the algorithm can be easily modified to solve mixed additive and multiplicative models as well as constraint-optimization problems (parameters that need to be capped) that actuaries often deal with in their practical work. Lastly, the algorithm is easy to understand and implement, and can be programmed with almost any software. A live example using Excel and Visual Basic for the algorithm will be used for the demonstration.

Moderator:

Panelists:

HOMEOWNERS

Download Handout

Many GLM ratemaking applications have focused on private passenger auto examples. This session will discuss how the nature of some homeowners’ variables affects a predictive modeling analysis. These include both traditional rating variables (such as amount of insurance, deductible, and policy form) as well as external variables related to demographics or weather. The typical indivisible premium approach for analyzing homeowners’ data does not lend itself well to proper investigation of these explanatory variables; therefore, the presentation will outline a case for modeling homeowners separately by peril. The panel will also survey the myriad ways various companies have incorporated this information into their rating plans, and discuss the advantages and disadvantages of various approaches.

Panelist:

Homeowners Insurance Scores/Disability Pricing and Dental Fraud Detection: Supervised and Unsupervised Learning

Download Handout

For some time now, scores using predictive modeling techniques have been used for evaluating credit and assigning tiers. But what other types of scores might be developed for use in rating a risk? The first part of this session we will examine some possibilities for the homeowners line of insurance.

The second half of the session will examine disability pricing and dental fraud detection which are two very different data mining applications. Disability pricing is an example of supervised learning, meaning that for each exposure in the historic dataset the actual claim cost is known. In contrast, dental fraud detection is an example of unsupervised learning, meaning that it is unknown which claims in the historic dataset were fraudulent.

The rate structure for disability insurance needs to be easily understood by management, regulators, sales force, and customers. This session will demonstrate how complex predictive modeling techniques can be applied to isolate the impact of various claim drivers on expected claim costs.

The dental insurance business is characterized by high-claim volumes and low-claim amounts. Fraud is suspected to be prevalent but the low-claim amounts make manual investigation relatively expensive. Mining the claim data would seem to be a good approach. However, it is not known which historic claims are fraudulent and which aren’t. This session will detail how unsupervised learning can be an effective approach to identifying suspicious claims.

Panelists:

HOW TO USE PREDICTIVE MODELING TO INVESTIGATE CLAIMS

Download Handout 1
Download Handout 2

For centuries, mathematical models have been constructed to represent quantitative relationships among variables, including predicting outcomes, given the presence of input variables. This session will cover the main ideas in some recently developed claim modeling approaches and discuss references for the technical details that are publicly available. The context will be predictive modeling for the decision to investigate claims for excessive medical treatment and fraud. Both supervised and unsupervised methods will be covered. Techniques will include:

Fuzzy logic and controllers
Regression-based scoring systems
The EM technique for filling missing data and profiling medical billing patterns
Tree-based methods, including CART, TREENET, and Random Forest

Examples from real auto claim data will be discussed.

Practical problems in implementing such techniques will be covered as well as the various applications of predictive modeling to the claims function. Emphasis will be placed on the personal and commercial lines of insurance. Some applications to be presented include estimating claim settlement values, estimating the impact of law changes on claim values, identifying potential fraudulent claims, and managing the claims process. In addition, an overview of the insurance claims fraud problem will emphasize claims processing and fraudulent and abusive claims detection.

Panelists:

OVERVIEW OF R

Download Handout 1
Download Handout 2 (.r)
Download Handout 3 (.r)
Download Handout 4 (.r)
Download Handout 5 (.r)

R is a freely available open source computer language designed for statistical computing. This session will start with an introduction to R, including where to get it, and how to use it. Next, it will illustrate some actuarial applications of R in the areas of predictive modeling for insurance pricing and reserving.

Panelists:

PRACTICAL ISSUES IN MODEL DESIGN

Download Handout 1
Download Handout 2

Textbook descriptions of model development are normally presented with rather straightforward examples. However, applying predictive modeling to insurance data presents a range of particular challenges. In this presentation some of the frequently encountered issues in developing a predictive model with insurance data will be discussed, including:

One of the commonly encountered insurance issues is a high proportion of missing data. Options for diagnosing the impact and adjusting a model when missing data will affect the model parameters will be presented.
An additional issue encountered in applying a GLM is a nonlinear pattern in a predictor variable. Options for addressing this will be presented, including the incorporation of a spline in the application of a predictor variable.
Interactions are often used to allow for dependencies between predictor variables. Different options for incorporating such dependencies across discrete variables, polynomials, and splines will be discussed.
It may not always be possible to extract information in the format best suited to model. One example of this is when a series of claim payments cannot be linked to a single claim event. We will show how the over-dispersed Poisson can be used in this example.

Moderator:

Panelists:

PREDICTIVE MODELING FOR SMALLER COMPANIES

Market leaders have embraced the idea of predictive modeling and have sufficient data to produce credible results. Small- to medium-sized companies may wonder whether predictive modeling can help them, given their smaller data volume. This session will discuss why predictive modeling has become more important for these insurers, and ways to address some of the unique issues they face when developing models – including data needs, competitive analyses, implementation, distribution, and regulatory aspects.

The session will also cover some of the results obtained when applying these techniques and some of the advantages smaller companies have when approaching the predictive modeling process.

Panelists:

PREDICTIVE MODELING LIFECYCLE

Download Handout

Modelers frequently focus on the modeling step, but there are many other important steps involved in a predictive modeling project. This session will discuss a framework for the successful development and implementation of a predictive model.

In this regard, we will describe each of the steps necessary for the completion of model:

Business problem understanding
Initial data understanding
Data preparation for modeling
Modeling
Evaluation of the model
Deployment of the model

The modeler is involved in each of these steps, but to create a successful project other business functions, including business management, underwriting, claims, field operations, marketing, and IT, need to be involved as well.

Panelists:

PRICING OPTIMIZATION I

Recent years have seen an increase in the sophistication of predictive modeling of claims and, more recently, policyholder retention and conversion rates. Often, however, relatively manual techniques are then used to derive the actual rates to apply in practice, failing to leverage the full potential of the underlying analyses.

This session will explore how sophisticated price optimization methods can be used to determine rates that best match an insurer’s strategic profit and growth objectives. The session will describe some of the technical approaches to price optimization, outline some potential pitfalls to avoid, and discuss real examples of how such methods can improve performance in practice.

Panelists:

PRICING OPTIMIZATION II

Download Handout

Are you doing a good job of aligning your business strategies with your predictive modeling initiatives and are you fully leveraging the power of predictive modeling across multiple business segments?

This session will introduce enterprise-wide predictive modeling and discuss ways insurance companies can implement an analytical framework that enables predictive modeling-based business decisions across three major operational areas—pricing, marketing/retention, and underwriting. Enterprise-wide predictive modeling also incorporates a 360° feedback approach to implementing, measuring and reassessing business decisions, resulting in a scaleable approach that can meet an insurance company’s short-term and long-term business objectives.

The session will also include discussions on how insurance companies can enhance decision making across these three operational areas and react quickly to market forces by implementing an enterprise-wide predictive modeling framework.

In today’s climate, pricing strategy and competitive forces are widely recognized as the most important challenges facing personal lines insurers that are seeking to grow or defend their market shares without sacrificing profitability.

Price optimization approaches, which balance the trade-off between profit and sales volume based on customer behavior and the competitive environment, have been successfully implemented in other service-led industries and are now being used by the financial services industry as a means of long-term value creation.

This session will look at some of the practical steps that insurance companies can take to develop market and customer behavior models, and to build this knowledge into their pricing strategies.

Panelists:

PROJECT MANAGEMENT FOR PREDICTIVE MODELS

Download Handout
Download Handout

The use of predictive modeling tools continues to expand within the property/casualty industry. The need to manage both the development and implementation of these complex tools is critical to accomplishing the goals for the business unit. This session will discuss the aspects of managing a complex project and the issues to consider for successful implementation.

Panelists:

TERRITORIAL ANALYSIS: PUTTING YOUR COMPANY ON THE MAP

Download Handout 1
Download Handout 2

Geographic location is one of the most widespread and established rating factors in the insurer’s rating algorithm, yet it is one of the more challenging factors to define. Historically, geographical risk classifications such as fire protection classes, rating territories, and zones have been defined based on physical surveys, engineering studies, and data analysis. Only the large companies had been able to deviate from the standard industry classifications.

Traditionally, rate factors for these risk classifications were determined using one-way analysis techniques, even though geographical risk tends to be highly correlated with other risk factors. Because of this correlation, it is imperative that locational rating factors be analyzed within the context of a multivariate framework. As location tends to be made up of a large number of dimensions, many of which have sparse data, special multivariate techniques are required.

This panel will discuss some techniques for determining definitions and rating factors based on the location of the risk and historical data. The discussion will include pros and cons of various options and diagnostics that can be used to validate the results.

Panelists:

TOOLS FOR MODEL DEPLOYMENT, PERFORMANCE METRIES, AND BUSINESS INTELLIGENCE

Download Handout 1
Download Handout 2

In Who Say’s Elephants Can’t Dance?, author Louis Gerstner Jr., the former CEO for IBM, recounts how IBM improved its decision-making capabilities, reacted more quickly to business drivers, and ultimately brought about one of the greatest corporate turnarounds in history.

The insurance industry can learn from IBM by shifting from reactive to proactive management and using advanced analytics to address business drivers that affect an insurer’s bottom line performance. These techniques can help a company make critical decisions sooner and faster, better align strategic goals based on quantitative metrics, and improve the organizational flow of information, and thus better meet market needs, customer service, and competitive pressures.

Predictive modeling is a major investment that requires on-going monitoring, measurement of business impact, and adjustments to maximize ROI. This session offers two different perspectives on the crucial issue of deploying a predictive model and addresses performance metrics and proximity modeling.

Moderator:

Panelists:

USE OF SCORING IN MARKETING

Download Audio (.mp3)

Download Handout

Customer life cycles offer many opportunities for improving profitability - prospecting, lead generation, risk segmentation and selection, acquisition, retention, cross sell, upsell, attrition, and win back. But it is not just the contractual relationship; there is also price elasticity for brand value to consider over multiple policy periods for the “life time,” We will discuss how to define a customer, and then how predictive models and operational implementation can improve your company’s profitability, both immediately and in the future.

In addition, the marketing department of an insurance organization seeks alignment with their underwriting department in terms of what attributes are associated with a “good” risk when pursuing new business. External predictors used in a company’s underwriting/pricing models can be leveraged to achieve better alignment. Potential new accounts can be scored and ranked based on the likelihood of passing through a company’s underwriting filter. The accounts can also be compiled reflecting the relevant distribution channel. The distribution channel factors considered include proximity, type of account, and referral information related to the targeted account.

Panelists:

VEHICLE SYMBOL DEVELOPMENT

The use of more refined vehicle rating plans in personal and commercial lines by several market innovators is beginning to cause companies to take a closer look at their own plans in order to make sure that their plans in conjunction with their more segmented rating plans are producing the most accurate rates. This session will discuss some of the plans being used by these insurers and discuss alternative vehicle rating systems that make better use of vehicle’s individual characteristics.

Panelists:

VISUALIZING PREDICTIVE MODELING RESULTS

Download Handout

The proper graphical presentation of a predictive model can be a critical diagnostic tool in the analysis of model results. Graphical presentation is also a key communication tool to individuals in an organization who do not have a detailed background in predictive modeling. In this session, the presenters will draw on their experience from a variety of predictive modeling projects in order to demonstrate a number of graphical presentation methodologies that they have found critical in proper diagnosis and presentation of model results. This will include techniques to understand key aspects of the data, identify and analyze predictor variables, and summarize key model results to senior management. Selected elements of the presentation will be in case study format.

Panelists:

WHAT TO DO WHEN YOU CAN’T USE CREDIT

Download Handout

Credit based insurance scores have become an important part of many insurer’s rating and underwriting plans, and can result in significant rate impacts. However, given the regulatory and public scrutiny of the use of credit information, there is a concern of what would happen if insurers could no longer use credit. This is actually the case in several jurisdictions, and insurers are coping with this reality. This session will discuss some of the ways insurers have dealt with not being able to use insurance score information, and also offer additional suggestions of how insurers might address this situation.

Panelists:

Site Search

2007 Predictive Modeling Seminar Handouts

General Session/Keynote Speaker

John F. Elder IV Chief Scientist, Elder Research, Inc.

Concurrent Sessions

ADVANCED MODELING TECHNIQUES

COMMERCIAL AUTO PREDICTIVE MODELING: THE SAME AND COMPLETELY DIFFERENT

COMMERCIAL LINES PREDICTIVE MODELING – BOP

COMMERCIAL LINES PREDICTIVE MODELING FOR WORKERS COMPENSATION

CREDIT SCORING UPDATE

DATA QUALITY—RAISING YOUR ACTUARIAL IQ (IMPROVING INFORMATION QUALITY)

DIMENSION REDUCTION

Estimating Personal Auto Loss Costs that Vary by Address/Household Averaging

FREE AND CHEAP DATA SOURCES

FREQUENCY AND SEVERITY MODELING

GLM I: INTRODUCTION TO GENERALIZED LINEAR MODELS

GLM II

GLM III

GLM Offset/Geneneralized Iteration Algorithms

HOMEOWNERS

Homeowners Insurance Scores/Disability Pricing and Dental Fraud Detection: Supervised and Unsupervised Learning

HOW TO USE PREDICTIVE MODELING TO INVESTIGATE CLAIMS

OVERVIEW OF R

PRACTICAL ISSUES IN MODEL DESIGN

PREDICTIVE MODELING FOR SMALLER COMPANIES

PREDICTIVE MODELING LIFECYCLE

PRICING OPTIMIZATION I

PRICING OPTIMIZATION II

PROJECT MANAGEMENT FOR PREDICTIVE MODELS

TERRITORIAL ANALYSIS: PUTTING YOUR COMPANY ON THE MAP

TOOLS FOR MODEL DEPLOYMENT, PERFORMANCE METRIES, AND BUSINESS INTELLIGENCE

USE OF SCORING IN MARKETING

VEHICLE SYMBOL DEVELOPMENT

VISUALIZING PREDICTIVE MODELING RESULTS

WHAT TO DO WHEN YOU CAN’T USE CREDIT

John F. Elder IV
Chief Scientist, Elder Research, Inc.