Claim risk prediction in the automobile insurance industry: a statistical data-mining approach

This paper considers the problem of assessing claim risk in the automobile insurance industry. A statistical data mining approach based on categorical data analysis is proposed. The most relevant features are searched for by either independence or conditional independence analysis. The latter aims at finding the so-called Markov Blanket of the “claim” variable, that is the minimal set of variables that renders the remaining variables superfluous for what concerns claim prediction. The proposed methodology was applied to an extensive data set provided by a primary Italian insurance company. The most relevant features turned out to be “risk class” and “fraction” (whether the premium is paid yearly or not). On a testing dataset, the predictor based on these two features performed better than classification trees. Copyright © 2007 IFAC Keywords: automobile insurance data, risk analysis data-mining, statistical methods, classification, prediction