Data misrepresentation detection for insurance underwriting fraud prevention

Premium fraud concerns data misrepresentation committed by an insurance customer with the intent to benefit from an unduly low premium at the underwriting of a policy. In this paper, we propose a novel approach for evaluating the risk of underwriting premium fraud at the time of application in the presence of potentially misrepresented self-reported information. The aim of the approach is to support insurance companies in identifying fraudulent applications and their decisions to underwrite insurance contract propositions. Likewise, it can be use to make straight-through processing (i.e. automated) underwriting systems more fraudproof, by e.g., triggering a validation on applications prone to misrepresentations. Our approach is based on conditional density estimates for a set of validated contracts. The proposed approach does not require historical fraud labels and can adapt to changes in pricing policy. Moreover, the approach can be used to detect outliers in addition to predicting underwriting fraud and is extended to multivariate self-reported data. We further demonstrate a link between Shapley values in common conditional expectation problems and conditional density estimations to make our approach explainable. We report a case study involving motor insurance underwriting, in which a driver's identity and driving record can be misrepresented to benefit from an unduly low premium; the results indicate the effectiveness of the proposed approach for detecting and preventing underwriting fraud.

Automated summary

This paper proposes a novel approach to evaluate the risk of underwriting premium fraud by using conditional density estimates. The approach supports insurance companies in identifying fraudulent applications and can adapt to changes in pricing policy. It can also detect outliers and predict underwriting fraud.