Fraud Analytics for Motor Insurance
To identify & predict patterns that typically relate to fraudulent claims.
A leading motor insurance firm wanted to identify fraudsters who use different modus operandi, dissimilar insurers and fake identities, in order to control the fraudulent claim count. Data quality and its availability was one of the key challenges, as there were 5 disparate sources (including Excel). Couple of other key objectives were to move from asset-based pricing to risk-based pricing and build a proactive prediction model to predict likely frauds right at the time of customer on-boarding.
We had collated data from varied sources and performed a data quality assessment. We had created business rules & hypothesis and proved them using methods like Exploratory Analysis, Test for Hypothesis etc. With the help of Ml and DL based predictive algorithms like Ensemble Trees, Extra Randomised Trees, LSVM, CNN, RNN etc, we had identified several relationships and links.
Network Graphs were chosen as the core of the solution in order to create relationships between entities. Out of network elements were identified as anomalies and sent across for further investigation.
With successful identification of key variables in a fraud scenario and by making relevant changes to the risk strategy, the client conceded an accuracy in the fraud prediction by 80% and more. Using the Fraud-risk dashboard, the client gets a consolidated view on the fraudulent claims. Predictive plausibility of fraud as a percentage of likelihood before on-boarding a customer had helped to investigate only “likely” cases of fraud and measure its prediction as a percentage of the entire data.