“Forecasting Workers Compensation Severities And Frequency Using The Kalman Filter” Frank Schmid and Jonathan Evans presented by Jonathan Evans, FCAS, MAAA Actuary CAS Seminar on Ratemaking Atlanta, GA March 8, 2007 Dr. Frank Schmid Senior Economist NCCI 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 1 Frank Schmid, director and senior economist in Actuarial and Economic Services at the National Council on Compensation Insurance, recently accepted a Hicks-Tinbergen Medal from the European Economic Association (EEA). The award was presented for the research paper, "Capital, Labor, and the Firm: A Study of German Codetermination," which he coauthored with Gary Gorton of the University of Pennsylvania prior to joining NCCI. The EEA recognized the research paper as the best paper published in the Journal of the European Economic Association in 2004 and 2005. 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 2 Forecasting Frequency And Severity Is Crucial To Workers Compensation Ratemaking • Prospective loss costs are very sensitive to trends in frequency and severity • Trend rates change over time • Forecasting changes in trend rates, or even turning points, greatly enhances rate adequacy 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 3 Forecasting As Signal Extraction And Extrapolation (NOT CURVE FITTING TO NOISE!) R2 = 55% R2 = 100% 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 4 Time Series Models • ARIMA - Auto Regressive Integrated Moving Average: focused on patterns of serial autocorrelation coefficients in observed data • UC – Unobserved Components: data assumed to be observed with white noise on top of signal • STS – Structural Time Series: combines UC with linear regression on exogenous explanatory time series 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 5 STS + UC Local Linear Model Observation (measurement) y t t t xt t, t ~ N (0, 2 ) Signal t t Exogenous Regression Parameter t t 1 t , t ~ N (0, 2 ) Level t t 1 t 1 t , t ~ N (0, 2 ) Slope t t 1 t , t ~ N (0, 2 ) The Local Level Model is the special case where the slope and exogenous regression parameter is set to constant 0. The Local Level STS Model is the special case where the slope is set to constant 0. 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 6 The Kalman Filter Uses estimates for σε, σν, ση, and σζ, together with actual observations of yt to filter out measurement noise εt and produce a piecewise least squares estimate θt , similar to Bϋhlmann credibility. Since the likelihood function for the observations has arguments εt and σε, the values of σε, σν, ση, and σζ, can be MLE estimated from the Kalman filter estimates for θt . ˆt yt yˆt ( yt , ˆ , ˆ , ˆ , ˆ ) n L t 1 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 1 ˆ ˆt2 exp 2 2 2ˆ 7 NCCI Frequency And Severity Applications • Objective to forecast the 3 year growth factor for the indemnity and medical severities, and frequency of claims (per on-leveled and wage adjusted premium) • 18 observed log growth rates for accident years 1986 through 2004 • Severity data on a paid basis • Models applied to log growth rates of data points – Local Level model used for severity log growth rates – STS Local Level model used for frequency with the change in unemployment as the exogenous explanatory series 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 8 Logarithmic Growth Rates of Indemnity and Medical Severities, State-Level Data, Accident Years 1987-2004 0.12 Indemnity Severity Medical Severity 0.10 L o g arith m ic R ate o f G ro w th 0.08 0.06 0.04 0.02 0.00 -0.02 -0.04 -0.06 -0.08 1985 1990 1995 2000 Accident Year 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 9 2005 2010 L o g R ate o f G ro w th (F req u en cy ) an d F irst D ifferen ce (U n em p lo y m en t) Logarithmic Growth Rate of Frequency and First Difference in Rate of Unemployment, State-Level Data, Accident Years 1987-2004 0.20 Frequency Rate of Unemployment 0.15 0.10 0.05 0.00 -0.05 -0.10 1985 1987 1989 1991 1993 1995 1997 Accident Year 1999 2001 2003 Note: The Rate of Unemployment was measured in percent; for scaling purposes, the first difference was divided by 10 (in this exhibition only). 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 10 2005 Regression Diagnostics (Local Level UC Model) for the Log Growth Rate of Medical Severity 1 Correlogram 2 QQ Plot (Versus Normal) 1 0 0 -1 1 10 2 3 Lag Length 4 5 -1.0 1.0 Cumulative Sum of Residuals 0 -0.5 0.0 0.5 1.0 1.5 Cumulative Sum of Squared Residuals 0.5 -10 1990 1995 2000 Accident Year 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 2005 11 1990 1995 2000 Accident Year 2005 Holdout-Window Forecasts (Local Level STS Model) for the Growth Rate of Frequency 0.075 Actual Forecasts 0.050 Logarithmic Rate of Growth 0.025 0.000 -0.025 -0.050 -0.075 -0.100 -0.125 -0.150 -0.175 2001 2002 2003 Accident Year 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 12 2004 2005 Logarithmic Rate of Growth Forecasts (Local Level UC Model) for the Log Growth Rate of Medical Severity Actual Forecasts 0.10 0.05 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2002 2003 2004 2005 2006 2007 2008 Logarithmic Rate of Growth Accident Year Level (Trend Log Growth Rate) 0.08 0.06 0.04 1995 1996 1997 1998 1999 2000 2001 Accident Year 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 13 Kalman Filtered Forecasts Versus Forecasts Disregarding Measurement Noise For the holdout forecast for medical severity presented: • Kalman filtered forecasts of the annual log rates of growth have a sum of absolute forecast error (for periods T+1, T+2, and T+3) equal to 0.0387, and RMSE (root mean squared error) of 0.0090 • For the last observed rates of growth, the absolute forecast error is 0.1154 and the RMSE is 0.0234 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 14 Conclusion • The experience of NCCI with Kalman filtered estimation of trend rates during the policy year 2006 rate filing season was encouraging • Current research at NCCI has shifted from Kalman Filter+MLE estimation to Bayesian estimation (Gibbs sampling using WinBUGS) of underlying models similar to the UC and STS models in the paper 2007 National Council on Compensation Insurance, Inc. All Rights Reserved. 15