An insurance company processes millions of claims, knowing that some of these might be fraudulent. Fraudulent claims that get paid out can cause insurance companies unnecessary and sometimes massive losses. One question that is often posed, and one that is of great interest to actuaries and claims departments alike, is – How can we detect that a claim is fraudulent before it is paid out?

Machine Learning methods can greatly help automate and streamline the fraud detection process. They can give us insights into data that are not necessarily otherwise obvious. Such methods can help detect anomalies in data and can further help understand the causes and key drivers behind these anomalies. Actuaries can use this information in order to build better pricing models and advise companies on how to prevent losses due to fraud.

This presentation, given at GIRO in September 2019, explores certain such methods, how they were used on a sample Motor Insurance dataset to classify a claim as a fraud or not, and conclusions about the data drawn from these methods.