A Test of Clinical Judgement Vs. Statistical Prediction in Loss Reserving for Commercial Auto Liability

This paper is a case study of the quality of clinical judgment in loss reserving for Commercial Auto Liability in the U.S. for accident years 1995 through 2001. Research on clinical vs. statistical prediction in non-insurance fields indicates that relatively simple models frequently produce better results than human experts with access to the same information. To test the quality of clinical judgment vs. statistical prediction in the Commercial Auto Liability loss reserving process, we compared the ultimate loss ratios actually booked by the U.S. insurance industry for these accident years at twelve, twenty-four and thirty-six months of development to comparable loss ratio estimates generated by mechanical application of several basic loss development methods. The booked ultimate loss ratios differed significantly from those indicated by the mechanical application of chain ladder and Bornhuetter Ferguson methods, implying that the booked ultimate loss ratios were not determined using those methods, at least not without significant adjustment. We then compared all of these booked and estimated loss ratios to the ultimate loss ratios booked as of the end of 2004, which we treated as proxies for the true ultimate loss ratios. In most cases, the mechanically generated ultimate loss ratio estimates were closer to the booked estimates as of the end of 2004 than were the earlier booked loss ratios. The conclusion must be that, either the booked ultimate loss ratios were based on other methods that are inferior to the chain ladder and Bornhuetter-Ferguson or judgmental adjustments were made to the indicated ultimate loss ratios that reduced the quality of the final selections. Further research would be required to determine whether this is a general loss reserving phenomenon or one confined to Commercial Auto Liability during the tame period studied.