Applying Machine Learning to Anomaly Detection in Car Insurance Sales

Abstract

Financial revenue, in the insurance sector, is systematically rising. This growth is, primarily, related to an increasing number of sold policies. While there exists a substantial body of work focused on discovering insurance fraud, e.g. related to car accidents, an open question remains, is it possible to capture incorrect data in the sales systems. Such erroneous data can result in financial losses. It may be caused by mistakes made by the sales person(s), but may be also a result of a fraud. In this work, research is focused on detecting anomalies in car insurance contracts. It is based on a dataset obtained from an actual insurance company, based in Poland. This dataset is thoroughly analysed, including preprocessing and feature selection. Next, a number of anomaly detection algorithms are applied to it, and their performance is compared. Specifically, clustering algorithms, dynamic classifier selection, and gradient boosted decision trees, are experimented with. Furthermore, the scenario where the size of the dataset is increasing is considered. It is shown that use of, broadly understood, machine learning has a realistic potential to facilitate anomaly detection, during insurance policy sales.