Insurers today have been under tremendous pressure to increase their overall profitability. With their topline growth restricted to a single digit, insurers are now looking at improving their bottom line by reducing the cost of risks and by managing other operating expenses. In order to achieve these objectives, it is imperative that all insurers look at claims leakages as one of the cost saving avenues - which can effectively contribute 5%-10% improvement to the combined ratio.
One of the major leakages in claims is due to fraudulent claims. According to Federal Bureau of Investigation & National Insurance Crime Bureau, in the US alone, the cost of non-health insurance fraud is about $40bn annually, which means, it contributes to $400 - $700 increase in the premium per household. "Apollo™", which is a Big Data platform solution, tackles these challenges of fraudulent claims by applying predictive machine learning algorithms. Pattern based approaches to detect fraudulent claims allows the model to adapt to the changing claim patterns and learn from investigator feedback.
The approach to detecting fraudulent claims using Wipro Apollo™ platform involves the following steps: Data preparation and feature engineering, iterative model and feature selection and finally, model prediction and re-calibration based on investigator feedback. For e.g. in auto insurance, a delay in notifying the police about an incident or delay in notifying the victim's state in the vehicle involved in the accident etc. can be features. Models are then built for each fraud type by iterating between machine learning algorithms, features and parameters. These models alleviate class imbalance between fraudulent and non-fraudulent claims using multiple techniques like SMOTEBoost and RUSBoost algorithm. The Multi-Variate Gaussian models that uses PCA for dimensionality reduction detects outlier claims that do not adapt to the model.
An illustrative distribution of the Multi-Variate Gaussian model for 2 dimensions (of the 14 dimensions after dimensionality reduction) is as below: Red dots indicate known frauds in the sample set while green are normal claims. The elliptical band is the threshold for flagging investigations which is set based on user preferences for the relative weightage of precision vs recall.