Combining Predictions of Auto Insurance Claims.

This paper aims at achieving better performance of prediction by combining candidate predictions, with the focus on the highly-skewed auto insurance claim cost data. We analyze a version of the Kangaroo Auto Insurance company data, and incorporate different combining methods under five measurements of prediction accuracy. The results show: 1) When there exists an outstanding prediction among the candidate predictions, the phenomenon of the 'forecast combination puzzle' may not exist. The simple average method may perform much worse than the more sophisticated combining methods; 2) The choice of the prediction accuracy measure is crucial in defining the best candidate prediction for 'low frequency and high severity' datasets. For example, Mean Squared Error (MSE) does not distinguish well different combining methods, the MSE values of which are very close to each other; 3) The performances of different combining methods can be drastically different. 4) Overall, the combining approach is valuable to improve the prediction accuracy for insurance claim cost.