Testing Models on Simulated Data Presented at the CAS Annual Meeting November 13, 2007 Glenn Meyers, FCAS, PhD ISO Innovative Analytics The Application Estimating Loss Reserves • Given a triangle of incremental paid losses – Ten years with 55 observations arranged by accident year and settlement lag • Estimate the distribution of the sum of the remaining 45 accident years/settlement lags • Loss reserve models typically have many parameters – Examples in this presentation – 9 parameters Danger – Possible Overfitting! • Model describes the sample, but not the population • Understates the range of results • The range is the goal! • Is overfitting a problem with loss reserve models? • If so, what do we do about it? Outline • Illustrate overfitting with a simple example – Example – fit a normal distribution with three observations – Illustrate graphically the effects overfitting • Illustrate overfitting with a loss reserve model – Reasonably good loss reserve model – Show similar graphical effects as normal example Normal Distribution • MLE for parameters m and s 1 n m ˆ xi and s ˆ n i 1 • n = 3 in these examples 1 n 2 xi mˆ n i 1 Simulation Testing Strategy • Select 3 observations at random Population - m = 1000, s = 500 • Predict a normal distribution using the maximum likelihood estimator • Select 1,000 additional observations at random from the same population • Compare distribution of additional observations with the predicted distribution Simulated Fits Simulation 24 7000 7000 9000 11000 Simulation 91 7000 9000 11000 Loss Loss Simulation 132 Simulation 4 9000 Loss 11000 7000 9000 Loss 11000 PP Plots 0.6 0.4 0.2 0.0 Predicted Percentiles If predicted percentiles are uniformly distributed, the plot should be a 45o line. 0.8 1.0 Plot Predicted Percentiles x Uniform Percentiles 0.0 0.2 0.4 0.6 Uniform Percentiles 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.4 0.8 Simulation 91 0.0 Predicted Percentiles of Outcomes 0.4 0.8 Simulation 24 0.0 Predicted Percentiles of Outcomes PP Plots 0.0 0.2 0.4 0.6 0.8 Uniform Percentiles 1.0 0.6 0.8 1.0 0.4 0.8 Simulation 4 0.0 Predicted Percentiles of Outcomes 0.4 0.8 Simulation 132 0.0 0.4 Uniform Percentiles 0.0 Predicted Percentiles of Outcomes Uniform Percentiles 0.2 0.0 0.2 0.4 0.6 0.8 Uniform Percentiles 1.0 Simulated Fits Simulation 24 7000 7000 9000 11000 Simulation 91 7000 9000 11000 Loss Loss Simulation 132 Simulation 4 9000 Loss 11000 7000 9000 Loss 11000 View Maximum Likelihood As an Estimation Strategy • If you estimate distributions by maximum likelihood repeatedly, how well do you do in the aggregate? • Consider a space of possible parameters for a model • Select parameters at random – Select a sample for estimation (training) – Select a sample for post-estimation (testing) Continuing Prior Slide • Select parameters at random – Select a sample for estimation (training) – Select a sample for post-estimation (testing) • Fit a model for each training sample • Calculate the predicted percentiles of the testing sample. • Combine for all samples. • In the aggregate, the predicted percentiles should be uniformly distributed. 1.0 PP Plots for Normal Distribution (n = 3) 0.6 0.4 0.0 0.2 Predicted Percentiles of Outcomes 0.8 ML Bias Corrected ML 0.0 0.2 0.4 0.6 0.8 1.0 Uniform Percentiles S-Shaped PP Plot - Tails are too light! Problem – How to Fit Distribution? • Proposed solution – Bayesian Analysis • Likelihood = Pr{Data|Model} • We need Pr{Model|Data} for each model in the prior Predictive Distribution Model Distribution Pr Model|Data Model Bayesian Fits • Predictive distributions are spread out more than the MLE’s. • On individual fits, they do not always match the testing data. 7000 Simulation 24 Simulation 91 Bayes ML Bayes ML 9000 11000 7000 11000 Loss Loss Simulation 132 Simulation 4 Bayes ML 7000 9000 9000 Loss Bayes ML 11000 7000 9000 Loss 11000 Bayesian Fits as a Strategy Combined PP Plot for 0.0 0.2 0.4 0.6 0.8 1.0 Bayesian Fitting Strategy Predicted Percentiles of Outcomes • Parameters of model were selected at random from the prior distribution • Near perfect uniform distribution of predicted percentiles • At least in this example, the Bayesian strategy does not overfit. 0.0 0.2 0.4 0.6 Uniform Percentiles 0.8 1.0 Analyze Overfitting in Loss Reserve Formulas • Many candidate formulas - Pick a good one E Paid LossAY ,Lag PremiumAY ELR Dev Lag • Paid LossAY,Lag ~ Collective Risk Model • Claim count distribution is negative binomial • Claim severity distribution is Pareto – Claim severity increases with settlement lag • Calculate likelihood using FFT Simulation Testing Strategy • Select triangles of data at random – Payment pattern at random – ELR at random – {Loss|Expected Loss} for each cell in the triangle from Collective Risk Model • Randomly select outcomes using the same payment pattern and ELR • Evaluate the Maximum Likelihood and Bayesian fitting methodology with PP plots. Background on Formula • Same formula appeared in SessionV2 • Fit a Bayesian model to over 100 insurers and produced an “acceptable” combined PP plot on test data from six years later. • This paper tests the approach to simulated data, rather than real data. 0.25 Prior Payout Patterns 0.00 0.05 0.10 D ev Lag 0.15 0.20 Fixed Subpopulation 2 4 6 L ag 8 10 Prior Probabilities for ELR ELR 0.600 0.625 0.650 Prior 3/24 4/24 5/24 ELR 0.675 0.700 0.725 Prior 4/24 3/24 2/24 ELR 0.750 0.775 0.800 Prior 1/24 1/24 1/24 Selected Individual Estimates Simulation 4 Simulation 6 Bayes ML 50000 70000 90000 110000 Bayes ML 50000 70000 90000 Loss Loss Simulation 12 Simulation 17 Bayes ML 50000 70000 90000 Loss 110000 110000 Bayes ML 50000 70000 90000 Loss 110000 0.6 0.4 0.2 0.0 Predicted Probability • PP plot reveals the S-shape that characterizes overfitting. • The tails are too light 0.8 1.0 Maximum Likelihood Fitting Methodology PP Plots for Combined Fits 0.0 0.2 0.4 0.6 Uniform Probability 0.8 1.0 0.6 0.4 0.2 0.0 Predicted Probability Nailed the Tails 0.8 1.0 Bayesian Fitting Methodology PP Plots for Combined Fits 0.0 0.2 0.4 0.6 Uniform Probability 0.8 1.0 Summary • Examples illustrate the effect of overfitting • Bayesian approach provides a solution • These examples are based on simulated data, with the advantage that the “prior” is known. • Previous paper extracted prior distributions from maximum likelihood estimates of similar claims of other insurers Conclusion • It is not enough to know if assumptions are correct. • To avoid the light tails that arise from overfitting, one has to get information that is: Outside The Triangle