Insurance Fraud Prediction with Machine Learning

Insurance Fraud is a problem which has received little attention in the machine learning literature. The two common fraud types encountered in insurance are Claims and Underwriting fraud (Viaene, 2004). One specific feature shared by both types is that the fraud status in non-self- revealing, which challenges the completeness and existence of a label history. This finding challenges the usage of machine learning approaches, which have been successfully applied to other fraud types (notably Credit Card Fraud). At underwriting of an insurance contract, fraud is as dynamic as the business environment itself and prospective policyholders swiftly capitalize on the latest opportunities which are set by the pricing policy of the company. Little recent research has been conducted on the topic (Xia, Gustafson, 2016; Akakpo et al. 2019), using restrictive assumptions (parametric distributions, unidirectional misrepresentation). A quantitative approach to model adverse data misrepresentations with respect to a pricing model using state of art approaches in non-parametric statistics is proposed. At the occurrence of a claim, the usual independence of records -a common assumption in supervised and unsupervised fraud problems- is often violated. Fraudsters organise staged accidents and construct social networks to perpetrate fraud. Graph networks capture those interactions, however their structures are not suited to common machine learning approaches which expect tabular data. This research aims at investigating the most effective way to approach fraud problems in insurance specifically, both at claim and underwriting stages, in collaboration with Allianz Benelux.