Detecting Fraud In
Motor Insurance Claims
Dispatching Surveillance on Suspects
In 2016, a German insurance company serving the Southeast Asian market consulted PulseMetrics to develop its capabilities for fraud detection on its motor insurance claims. From on an exploratory data analysis, our consultants showed that though our clients’ insurance policy claims ratio stood at 30 percent, the insurance claims loss ratio was more than 50 percent. This meant that only 30 percent of the customer base and the insured vehicle population had claims but the insurance company was losing half of what it earned on its premiums. It proved that our client was incurring a higher than expected loss from its customers’ insurance claims. Therefore we used advanced analytics to predict if some of the claims were fraudulent, in order to minimise these unwarranted losses.
Building Up Investigative Capabilities
Our consultants commenced the project by creating scenarios where fraud was possible during the customer journey, such as the time of claim relative to the tenure or collusions with workshops on the prices of spare vehicle parts and the labour cost.
We then designed a heuristic scoring model for the likelihood of fraud, which is based on three supplementary scores of varying weightage: customer score, loss assessor score and workshop score. First, we conducted thorough analyses of the data on the tenures, policies, claims and vehicles, to rank customers on the likelihood of being fraudulent. Then we designed a decision tree model to predict the probability of customers having a fraudulent claim, and cross-referenced it with the enriched customer profile attributes such as age, gender, vehicle type, and distance from workshops to better understand the profile of customers who had the highest fraud likelihood score. For the loss assessor, we looked at the distribution and variance in their estimated costs of spare vehicle parts and labour cost. Lastly, the workshop score was based on a comparatively analysis of spare vehicle parts and labour costs among similar claims by different workshops. Using clustering algorithms and regression analyses, we identified the outliers and calculated the score based on the difference between the market average and the predicted costs by the workshops.
Searching for the Smoking Gun
The combined fraud likelihood score for insurance claims presents irrefutable evidence that collusion has taken place. Looking deeper at the components, a high loss assessor score or workshop score will point to who the suspect is. Through our analytical work, we have thus enabled the insurance company do targeted investigations to find out if the loss assessor and/or the workshop has committed fraud. This has helped our client weed out the unprofitable customers and channels, thereby hacking the customer journey post-acquisition to ensure that claims are made through legitimate means.
Financial Services (Insurance)
Surveillance & Risk Management
Decision Tree Modelling