Data Mining Approaches to Modeling Insurance Risk. Dan Steinberg, Mikhail Golovnya, Scott Cardell. Salford Systems 2009

Transcription

1 Data Mining Approaches to Modeling Insurance Risk Dan Steinberg, Mikhail Golovnya, Scott Cardell Salford Systems 2009

2 Overview of Topics Covered Examples in the Insurance Industry Predicting at the outset of a claim the likelihood of the claim becoming serious (CART example) Developing a model of total projected customer value for a health insurer Premium increase optimization Combination with GLM Other examples TreeNet tutorial

3 Predicting at the outset of a claim the likelihood of the claim becoming serious Consulting Project in Workers Compensation Insurance (with Pricewaterhouse Coopers Australia). Used Salford Systems CART. In worker s compensation insurance, serious claims comprise a small proportion of all claims by number but the great majority of the incurred cost. In our case study, 14% of reported claims were classified as serious. Those claims made up around 90% of total claim cost. Not obvious, in most cases, which claims will become serious as there are many factors contributing to result.

4 Predicting at the outset of a claim the likelihood of the claim becoming serious (continuation) Results: Data Mining (CART) identified 19 variables as most important predictor variables from 83 variables, including categorical variables with 100 s of categories. Classified all claims as likely to become serious or not likely to become serious. 31% of claims classified as serious Of these, 10% turn out to be serious 21% turn out to be false positives 69% of claims classified as non-serious Of these, 4% subsequently prove to be serious Interesting: Expected Result: Injury details were important in predicting serious claims Unexpected Result: language skills of the claimant were important in predicting serious claims Copyright Salford Systems 2009

5 Predicting at the outset of a claim the likelihood of the claim becoming serious (continuation) Gains Chart: 40% lift First 20% of the population, 60% of serious claims.

6 Developing a model of total projected customer value for a health insurer Consulting Project in Health Insurance (with Pricewaterhouse Coopers Australia). Used Salford Systems CART and MARS. Lifetime customer value is the discounted present value of income less expenses associated with a customer

7 Developing a model of total projected customer value for a health insurer (continuation) Result: Predicted the high cost claimants with good degree of accuracy. 15% of members predicted as having the highest cost by the data mining model, yielded 56% of the total actual cost. 30% of the members predicted as having the highest cost by the data mining model, yielded 80% of the total actual cost. Much of the model is accounted for by the dependence of cost on age. The model improves the gains chart by around 10% of total cost in the high cost part of the population over a simple model depending on age. Details of the resulting model are commercially sensitive. Results showed that many other factors, in addition to age, contribute to the predicted cost.

8 Developing a model of total projected customer value for a health insurer (continuation) How CART was used: Exploratory analysis: preliminary tree model suggested segmenting the customer base into 4 broad groups according to age and previous claims experience Separate CART model created for each of the 4 broad segments discovered in Exploratory Analysis phase: Sometimes risk drivers were similar but in many cases there were significant differences How MARS was used: Included CART decision tree segment for each record in the form of a categorical variable as one of the input variables. Also included as inputs: variables selected by CART as important predictors. MARS ranked variables in order of their importance. The predictor variable representing the CART decision tree output was ranked as most important Other, mostly continuous variables were also ranked as important which showed that the MARS model was finding minor linear effects that were not picked up by the decision tree. Copyright Salford Systems 2009

9 Premium Increase Optimization, TreeNet/GLM example Consulting Project in Home Insurance (with Taylor Fry Consulting Actuaries) Used Salford Systems TreeNet and GLM Although Home Insurance example, also applicable to Health Insurance Problem: Average renewal rates are around 87.5% Current renewal price is risk cost + expenses + target profit Constraint: +/- $50 from current price Details camouflaged at customer request Goal: pricing insurance products according to what customer is prepared to pay What is best that can be achieved? Insurance profit vs renewal rate

10 Premium Increase Optimization, TreeNet/GLM example (continuation) Approach Taken First Modeling Step: Used TreeNet: Quickly provided a benchmark model Variable selection Feature identification 2 nd Step: Built GLM Model 3 rd Step: Assessed Performance of GLM model versus Benchmark model Without data mining techniques, model building would take far longer. TreeNet model gives performance that is usually close to (sometimes better than) the final model adopted for demand modeling.

11 Premium Increase Optimization, TreeNet/GLM example (continuation) Results: Knowledge of drivers of renewal delivers quick wins Segment into high/low renewal Analyze drivers and levers Manage low expected renewal rate segments throughout the renewal process Identified customers less sensitive to price increases

12 Other Examples Contact Salford Systems: Actuarial Case Study Presentations Available Combining Linear and Non-Linear Modeling Techniques: Case Study Example from the Insurance Industry Insurance Fraud Detection Use of Data Mining (CART) and Rule-based Methods to Infer Claims Payment Policy from the Analysis of Paid Claims Text Mining Challenges and TreeNet: Impact of Textual Information in Claims Cost Prediction Statistical Case Estimation Modelling - An Overview of the NSW WorkCover Model CART Plus Linear Models for Actuarial Applications Insurance Premium Increase Optimization: Case Study And more.

13 Other Examples Contact Salford Systems: Marketing Case Study Presentations Available: Marketing Strategies for Retail Customers Based on Predictive Behavior Models Modeling Effectiveness of Marketing Campaigns How to Exploit the Homogeneity of Data or the Homogeneity of Past Experience in High Volume Direct Marketing Programs Predicting Customer Behavior Trends Over Space and Time CART Analysis to Augment Predefined Market Zones & Boost Response Rates for Direct Mail Campaigns High-Level Marketing Strategy: Deciding Whether to Customize Product Offerings and Promotion and Whether to Reward Best Customers with Perks Application of CART in Determining Reserve Levels for Customer Loyalty Points Achieving Better Insights Into Customer Behavior Through Integrating Market Research with Decision Tree Behavior Modeling

14 Interaction Detection with TreeNet Methodology Example

15 The challenge of interaction detection Classical statistical modeling is focused primarily on the development of linear additive modeling Models are of the form Y = A + B1X1 + B2X2 + +BkXk The predictors or attributes Xi are either raw data columns or data after repairs are made such as missing value imputations and capping or elimination of extreme values Non-linearity is introduced into the models in limited ways such as via the log transform of Y (for continuous Y>0) and a collection of well known transforms of the Xi Credit risk scorecard technology introduces transforms derived from binning continuous predictors

16 Classical Statistical Model Performance Classical statistical models represent a huge fraction of real world models deployed in enterprises world wide Such models are popular in part because statisticians are well trained to develop them The models are also popular because they tend to give good performance no matter how such performance is measured (e.g. R-squared, Area Under the ROC curve, top decile lift, etc) Such models are so restrictive they must over-simplify but they benefit from their stability (low variance)

17 Bias/Variance Trade-Off The universe of possible models is huge whereas the subset of linear or linear models is quite small Forcing a model to conform to a specific mathematical form is likely to introduce distortions We are therefore willing to say that the classical models are almost certainly biased As data become more plentiful the models almost certainly converge to an incorrect representation of the data generating process In contrast, modern learning machines such as TreeNet and CART converge to a correct unbiased representation Classical models have the advantage of lower variance which might translate into an overall better model in smaller samples

18 Classical Model Shortcomings: Sources of Bias The classical statistical model is expected to fall short in two areas: Incorrect representation of nonlinearities in individual predictors Absence of relevant interactions Interactions will be the key to identifying important data segments which behave differently from the dominant patterns Specifying a model correctly is the most challenging task facing the modeler Until now there has been no reliable way for a modeler to determine whether they have found a correct specification

19 Classical Interaction Detection A popular method of statistical interaction detection is to begin by first building the best possible additive model. Then, interactions terms, which look like products of main effects, e.g. Xi*Xj are added to the model and tested for significance One problem with this method is that the number of possible interactions grows more rapidly than square of the number of predictors All the usual challenges of model construction apply: the XiXj may not become visible unless one also includes XmXn There may well be higher order interactions XiXjXk Interactions may only exist and may only be detectable in subregions of the predictors (Xi Xi>c)*(Xj Xj<d)

20 What is Needed The classical approach of building the best possible additive model and then testing for interactions is largely impractical and largely infeasible What is needed is a methodology that can automatically discover the right transformations of each predictor and introduce the right interactions This methodology must be far more flexible than classical modeling and yet be resistant to overfitting It must be able to tell us whether or not interactions are present and to identify precisely which interactions are present

21 TreeNet and Model Development TreeNet is Jerome Friedman s stochastic gradient boosting first described in Starting from the base of MART (Multiple Additive Regression Trees) TreeNet has evolved over the past decade into a powerful modeling machine. TreeNet can fit a variety of models, including regression and logistic regression. The models are non-parametric and based on hundreds if not thousands of small regression trees. Each tree is designed to learn very little from the data and thus many are needed to complete the learning process Models are in the form of error correcting updates

22 Some interesting TreeNet Features Trees are grown on only a random subset of the training data, typically a random half We never train on all the training data at one time Model updates are small. We do not allow the model to change more than a little in any training cycle The equivalent of outlying data (points that are badly mispredicted) are eventually ignored in training cycles We do not allow anomalies to have much influence on the model

23 A TreeNet Model For the binary dependent variable it can be shown that the TreeNet model is a non-parametric logistic regression For the continuous dependent variable the TreeNet model is a nonparametric regression fit to maximize (or minimize) one of the following objective functions: Least Squares Residuals Least Absolute Deviations Huber-M hybrid of LS and LAD (LS for small residuals, LAD for large)

24 Some Key TreeNet Features Automatic variable selection. TreeNet is a superb ranking machine. TreeNet tends to include more predictors than other learning machines TreeNet tends to give effective ranking of predictors allowing the modeler to select small subsets of reliable predictors TreeNet contains built-in step-wise variable selection (backward, forward, or backward/forward) to automatically search for a best model Start with all variables in model Using the variable importance ranking remove the R least important predictors (typically R=1) Repeat until all predictors have been removed Identify best model based on preferred performance metric

25 Forward Stepping Starting with best model test every predictor in a specified subset to see which is best to add Repeat forward stepping for F steps Identify best new model which is a final model candidate

26 Testing for Interactions When the TreeNet model is built with only 2-node trees the model is essentially limited to an additive model With a single split in the tree the tree cannot capture any interactions A 2-node tree TreeNet is thus an additive model. Observe that an additive model can be highly nonlinear Model is of the form Y = A + F1(X1) + F2(X2) +..Fk(Xk) Where the Fi are the nonlinear functions discovered by TreeNet Note that there may be many Fi associated with a specific predictor Xj ( cumulate these into G(Xi) = Σ Fq(Xi) ) If the TreeNet contains only 2-node trees we can collect all trees associated with a given Xi to arrive at the final Y = A + G1(X1) + G2(X2) + + Gk(Xk) Copyright Salford Systems 2009

27 Building 2-Node TreeNets Important to keep in mind that when only one split is allowed in a tree the amount of learning that can take place is severely limited. 2-node tree TreeNets may require a very large number of trees to extract all the information in the data Some examples we will show use 20,000 trees Unless the TreeNets are fully expanded it will not be possible to measure the true predictive power of the additive model

28 General TreeNet Models TreeNet models are permitted to contain trees of any size We can grow trees with 2,3,4,5,6,12,50 etc nodes We generally favor small trees because we want to limit the amount of learning in any training cycle We therefore tend to keep trees to sizes like 6,9,12. However experimentation is always recommended and we have encountered data where larger trees perform better

29 The Default 6-Node Tree Friedman recommended a default setting of 6-nodes and our experiments confirm that this size of tree performs well across broad range of data sets A 6-node tree clearly permits interactions. A tree with 6 terminal nodes contains 5 internal splits and if it is as close to balanced as possible like the tree below then it can contain up to 3 different variables along its longest branch If the tree is maximally unbalanced then it can contain up to 5 different variables along the longest branch There is no guarantee that different variables will be used as we progress down a tree. The same variable might be used several times. In general we observe that the 6 node tree should be adequate to uncover 3-way interactions

30 Global Intercation Test Compare Unrestricted TreeNet model allowing moderate sized trees Restricted TreeNet model confined to just 2-node trees Must take into account the total learning in each model (number of trees * number of nodes per tree) Otherwise the 2-node tree model will be at an automatic disadvantage Simply allow each model to reach convergence

31 A Global Interaction Test Consider a TreeNet model built with 2-nodes and compare this model with a 6-node TreeNet model Recall that the TreeNets must have been allowed to grow out fully to locate the optimal number of trees for prediction By comparing the restricted 2-node tree model with the 6- node tree model we can conduct a definitive test for the presence of interactions of moderate degree: If the larger tree (unrestricted) model sufficiently outperforms of the (restricted) 2-node model then we have compelling evidence that interactions are present in the data generation process At present we do not have a definitive statistic for testing this hypothesis but classical statistical tests can be developed when comparing predictions of constrained and unconstrained models on holdout data

32 Which Interactions Are Present Our global test is sufficient to establish the existence of interactions but not sufficient to identify which specific interactions are present All we can conclude from a positive result of our test is that interactions exist and from a negative result that interactions do not exist However this is a major step forward as this test is fully automatic We have found evidence that many mainstream consumer risk models are adequately modeled with additive TNs The next step for us is therefore to try to identify specific interactions

33 Interaction Detection In TreeNet Interaction in TreeNet has progressed along two fronts. Friedman suggested one strategy in his paper. At Salford Systems, Cardell, Golovnya, and Steinberg suggested a slightly different method.

34 Interaction Measurement From the TreeNet model extract the function (or smooth) Y(Xi, Xj Z) (1) which is based on averaging the Y associated with all observed Xi, Xj pairs over all observed Z Now repeat the process for the single dependencies Y(Xi Xj, Z) and Y(Xj Xi, Z) (2) Compare the predictions derived from (1) and (2) across an appropriate region of the (Xi, Xj) space

35 A Simple Example: Financial Market Behavior Top 15% of values Bottom 85% of values Dependent Variable: Continuous with a long right hand tail

36 Summary Stats: Target and 9 Predictors TARGET variable BPS is truly continuous; others are ideal for a tree model We randomly set aside 20% of the data for test

37 Baseline CART Model: Test Data MSE (Random 20%) 90 node tree SE node tree SE R2 better than.97 on test data Observe that a CART model allows for high order interactions

38 Naïve Regression: R2 is only.74 on train data R2 on test data is.68

39 MARS Models results Main Effects (nonlinear) way interactions way interactions way interactions MARS main effects and interactions determined by 4-way model

40 TreeNet Results: Test MSE TreeNet 6-node unconstrained (3303 trees) TreeNet 2-node tree (20,000 trees) Although the starting sum of squares in the data is large it seems intuitive that the difference between the two models is substantial The 2-node TreeNet is similar in performance to the MARS model with 2-way interactions even though it prevents interactions The 2-node TreeNet underperforms a single CART tree

41 TreeNet Interactions Ranking Report Based on the 6-node tree TreeNet we calculate the degree of interaction observed for the most important variables This report reveals that e.g. AVG_VOL is involved in important interactions

42 Detailed Interaction Reports Measured interactions based on the comparison of 2D and 3D relationships between target and predictors Allows us to hypothesize which interactions are likely to matter

43 Nine predictors permit 36 2-way Top rated 2-way interactions (6 in all) interactions AVG_VOL * AVG_TREND * SPREAD_Q SPREAD_Q * AVG_TREND * AVG_SIZE AVG_TREND * AVG_SIZE AVG_SIZE * MKT_CAP

44 Testing Interactions Salford Systems has developed an Interaction Control Language (ICL) for TreeNet models The language allows the modeler to specify precisely the types of interactions which will be permitted in the TreeNet ICL allow X1 X2 X3 X4 / 2 Specifies that only 2-way interactions are allowed among the collection of predictors listed (X1-X4) The ICL language was developed in-house for private clients in 2006 and has been the basis of all of our interaction detection work

45 The ICL Language The ICL allows a broad range range of controls such as: ICL ADDITIVE x1 x2 x3 ICL ALLOW x5 x6 x7 x8 x9 / 3 ICL DISALLOW x9 x11 x5 x7 x21 x25 / 4 The ADDITIVE keyword prevents any predictor from interacting with any other variable in the model. Practically this means that should such a predictor be selected to split the root node of the tree than it can be the only predictor anywhere in that tree TreeNets restricted to ADDITIVE predictors can still contain trees with as many terminal nodes as the modeler prefers to work with. But each tree will be grown using a single predictor

46 ADDITIVE vs 2-node Trees 2-node trees are often thought to guarantee ADDITIVE models but this is not strictly true If the training data are complete (no missing values present) then indeed a 2-node tree TreeNet yields an additive (in the predictors) model However, if missing values are present then TreeNet requires the use of missing value indicator predictors of the form If Xi == MISSING then go LEFT; Else if CONDITION then go RIGHT; The problem is that TreeNet permits the condition in the ELSE clause to involve a variable other than Xi

47 2-node trees in TreeNet Although the user may request 2-node trees in a TreeNet model if missing values are present for all predictors then the smallest possible tree that can be grown contains at least 3-nodes. Here is an example of such a split

48 Two-Node Tree with Missings One variable in tree Is Xi missing? Yes No Terminal Node Is Xi <= c Terminal Node Terminal Node Both internal nodes are split using the same variable Xi

49 Two-Node Tree with Missings Two variables in tree Is Xi missing? Yes No Terminal Node Is Xk <= c Terminal Node Terminal Node Xi (root node split) and Xk (internal split) are the two variables in the tree

50 2 node tree details A 2-node tree in TreeNet can never contain more than one split on a standard predictor. Therefore if the predictor is never missing the tree using this variable will have only two nodes However, the TreeNet mechanism does not count a split on a missing value indicator as a genuine split As a result 2-node tree may contain any number of missing value indicator splits. The following tree is technically a 2- node tree by TreeNet standards.

51 Allowable 2-node tree Missing? Terminal Node Missing Terminal Node Split Terminal Node Terminal Node

52 2-node trees and Interactions The important point in this discussion is that in standard TreeNet we cannot guarantee the complete absence of interactions with 2-node trees The 2-node tree will allow interactions of any degree between missing value indicators and also a general interaction between a single predictor and missing value indicators To enforce literal non-interactivity we must rely on the ICL mechanism. If we specify that Xi is to enter the TreeNet as ADDITIVE then if we split the root using the MVI for Xi any subsequent split can use only Xi

53 Test MSE Performance: Restricted Models 2-node trees (20,000 trees) As expected the interaction report shows 0.00 for all interaction scores. The data contain no missing values so literal 2-node trees are generated for this model We require 20,000 trees because 2-node trees can learn only little in any training cycle

54 Test MSE Performance: Various Interactions Allowed 2-node trees (20,000 trees) Allow 2-way interactions Allow 3-way interactions Allow 4-way interactions Allow 5-way interactions Allow 6-way interactions Allow 7-way interactions Unconstrained TreeNet It is plain that there is a huge difference between an additive and an interaction model

55 Refining the Model Our choices include accepting a 2-way interaction model and searching for an explicit set of interactions to work with AVG_DLY_VOL, AVG_TRND, AVG_SIZE / MKT_CAP, AVG_DLY_VOL, AVG_TRND, AVG_SIZE, SPREAD_Q / Here we see that just 3 two-way interactions can capture more than 90% of the difference between an additive and a fully unconstrained model

56 TreeNet with ICL TreeNet in the PRO EX version contains the ICL language and is available on request from Salford Systems. Contact: David Tolliver General information

57 Salford Systems Developer of CART, MARS, TreeNet and RandomForests Advanced statistical software since 1983 PROC MLOGIT and PROC MPROBIT add-ins for SAS mainfames Technical Advisers and close collaborators Jerome H. Friedman, Stanford University ( CART, MARS, Treenet) Leo Breiman*, UC Berkeley (RandomForests) Richard Olshen, Stanford University (Survival CART, Bioinformatics) Charles Stone, UC Berkeley (CART, MARS large sample theory) Rob Tibshirani, Stanford (modern statistical methods) * Leo Breiman passed away in July 2005