Speaker(s): Jonathan Karsenty (PwC)
Given the current economic context, insurance fraud is a growing issue for all insurance companies and a day-to-day problem for claim handlers. As a proof, institutions are created all around the world to counteract fraud: UK insurance fraud taskforce, IFB (US insurance fraud bureau), ALFA (French insurance fraud bureau), et caetera. Detecting fraud using algorithms is a way to counteract fraud but accurate fraud detection in the insurance industry is hampered by a lack of data quality. Insurance fraud is not proven until an audit of the claim is made and most insurers would only audit claims under strong suspicion of their claim handlers. Data stored into insurers' databases are not completely reliable as they are conditional upon claim handlers' fraud suspicion.
To overcome this issue, this approach develops non-supervised learning methods: RIDIT and PRIDIT methods. Start by using the RIDIT method to calculate a fraud suspicion score for each variable of a dataset of claims. As RIDIT only applies to categorical variables, extend this method to continuous variables (without loss of accuracy). After applying RIDIT scoring for each variable, use PRIDIT method to obtain an overall fraud suspicion score for each claim.
According to the score, classify each incoming claim as "high fraud suspicion" or "low fraud suspicion" and help claim handlers on their decision as to start an audit. By this approach, practical, innovative and efficient methods to detect insurance fraud are developed. It allows to improve the fraud detection system already present in claims management and aims to automatize fraud detection and thus make the process of claim management faster.