Stochastic Reserving Challenge at the CLRS
Mark Shapland and Robert Bear
Did you miss your chance to participate in the stochastic reserving challenge? Everyone registered for the 2008 Casualty Loss Reserve Seminar in Washington, DC, was eligible to participate in the first stochastic reserving challenge. Perhaps because stochastic reserving is not yet commonplace, only 11 colleagues rose to the challenge. The results still proved quite useful, however.
The data set used in the challenge was realistic, but it was simulated data. Using the prototype CAS Loss Development Simulator (CASLDS) model created by the CAS Loss Simulation Model Working Party, we created parameters used to simulate individual claims that are then summarized into claim rectangles for 10 complete years. Using the parameters for individual claims, we then ran the model for 1,000 iterations so that we could determine the “actual” distribution of the claims being analyzed. While there were no exposure changes or inflationary trends in the simulations, we also randomly “adjusted” the earned premium around $3 million per quarter in order to create apparent fluctuations in the loss ratios.
Even though there were 1,000 data sets to choose from, we selected a data set for which the paid and incurred chain ladder projections were both close to the mean of the distribution, but the challenge was to estimate the standard error, 75th, 90th and 99th percentiles in addition to the mean. For the 11 sets of results, we did have two outliers (one person estimated ultimates and another estimated only IBNR, instead of the total unpaid) but most were close to the actual mean. Interestingly, all participants underestimated the standard error of the distribution, but the winners tended to get closest to the actual standard error.
The top three contestants were:
1st Place—Jeffory C. Schwandt, Regnier Consulting Group Inc.
2nd Place—Jessica (Weng Kah) Leong, Milliman, Inc.
3rd Place—Spencer M. Gluck, Guy Carpenter & Co. LLC
The winners received nice certificates to memorialize their “victory” and gift certificates for $100, $50, and $25, respectively. We would also like to thank Ian Asplund, David Clark, Dean Dorman, Paul Herzog, Glenn Hiltpold, Glenn Meyers, Alan Putney, and Richard Quitano, who all deserve an honorable mention for participating.
While the simulation model provided the basis for this challenge, it is our goal to use it for future research regarding the strengths and weaknesses of different stochastic models. If you would like more information on the challenge, you can visit the CAS Web Site. The CASLDS model can be downloaded. The working party is in the process of working with a consultant to create a new loss simulation model in an open-source format that should improve our ability to create simulated data sets, conduct research, and improve the model.