Development of an Expert System for the Automatic Detection of Automobile Insurance Fraud

The goal of this study is to develop a tool to aid insurance company adjusters in their decision making and to ensure that they are better equipped to fight fraud. This tool is based on the systematic use of fraud indicators. We first propose a procedure to isolate those indicators which are most significant in predicting the probability that a claim may be fraudulent. We applied the procedure to data collected in the Dionne-Belhadji study (1996). The model allowed us to observe that 18 of the 50 indicators used were significant in predicting the probability of fraud. Our study also discusses the model's accuracy and detection capability. The detection rates obtained by the adjusters who participated in the study constitute the reference point of this discussion. As shown in Caron-Dionne (1997), there is the possibility that these rates underestimate the level of fraud.Our second step was to develop software allowing us to use the results of the statistical model to estimate the probability of fraud in files and to decide whether or not an in-depth investigation should be conducted. This software contains the mathematical equation and the parameters calculated by the Probit model. As indicated in the report, these parameters reflect the data from all the firms having participated in the study and not from any one company in particular. It is not obvious that the same indicators would be significant or even that the coefficients would be the same for any one insurer in particular. Any insurer wishing to use the software is advised to carry out a systematic study of the company's own files.Once adapted to the insurer's use, the software can easily be used by claims adjusters. It would then be a matter of entering the indicators present in the files. The software will calculate the probability of fraud in a file and help the adjuster to decide whether an in-depth investigation is warranted