Fraud analytics solution for an NBFC client

An average vehicle claim in the UK costs £10k to £12k. For a company that receives 100 fraudulent claims every month, the 3% improvement in fraud detection rate could mean a £30k savings on claim sanctions.

About client

Our client is a non-banking insurance company offering a diverse portfolio, covering vehicle, home, and business insurance products. They customize these products to suit the needs of every customer. With a huge number of clients across the nation, they process tons of disbursements and claims every day. In order to have a seamless risk management process and increase customer satisfaction, they sought help from advanced analytics and machine learning.

Challenges

They had a basic framework to evaluate claims-related fraud. Following are some of the challenges their teams encountered with the current process.

Rules based on prior experience: Their current risk management process involves subjecting a handful of selected claims to risk sampling scrutinization. They had a set of rules developed to identify and flag potential fraud cases out of these selected claims. The above rules and fraud parameters are based purely on their prior experience with little or no modifications done over a while. 

Manual sampling: Processing claim verification manually was time-consuming for their teams, causing delays and putting customer satisfaction at stake. 

Besides, there increasing amount of transactional and behavioral data demanded hours of time to review manually. The review team had to follow a complex scrutinization where even slight negligence led to errors and inefficiencies that often went undetected. Example: deliberate deception of vehicle/property theft by consumers, billing malpractices, etc. As they were unable to detect fraudulent claims with high accuracy, their fraud detection rate was much less than industry standards.

Changing fraud patterns: Fraudsters have evolved a lot and change their patterns every time they submit a suspicious claim. This makes it impossible to follow a defined set of rules for risk identification and mitigation. Even though the client’s team was aware of these patterns, they could neither anticipate nor develop new sets of rules to prevent them in the future. Some of these patterns are too hard to identify with human intelligence alone, without spending considerable time analyzing them. 

False positive cases: When suspicious patterns were noticed, it was sent to the attention of their triage team. They had to do the heavy lifting and analyze multiple data points to see if the claim was indeed a fraud case or not. They had to be very careful with their final verdict as false positive cases can lead to unnecessary investigations or delayed settlements - all of which can trigger unrepairable reputational damage.

The solution

Our data science team developed a fraud analytics solution that exhibited a 3% improvement in fraud detection rate, saving the company a great deal of money and time wasted reviewing & approving fraudulent and other claims.

Built an ensemble of machine-learning models: We built an ensemble of machine learning models using multiple algorithms like random forest. The ensemble model increased accuracy while detecting anomalies in claim applications. Upon studying previously flagged cases and performing pattern analysis, we uncovered more rules and patterns for training the model.

Fast, seamless, and accurate claim processing: The model analyzes incoming claims and their supporting documents. Any item it identifies as fraud is sent to a financial/triage analyst. Human expertise and real-world experience are applied to approve or deny the flagged claim.

The accuracy rates improve as the model learns from the new data and the validations from financial analysts.

Stability and scalability: After being used for a small batch of data, the model has significantly improved. Now it’s able to handle real-time claims data and process them faster without employing multiple resources. The triage team is free from a heavy chunk of documentation and nuanced analysis and only has to review the flagged ones. 

Customer satisfaction: They receive claims at unpredictable rates. Despite that, they were able to process settlements on time—which significantly improved customer satisfaction. Not having to go through additional investigations or higher wait times struck a chord with them.

Conclusion

Modern problems require innovative solutions. Financial fraud is one such, causing institutions to come up with rigid measures and rule-based identification systems. Yet, the fraud rate doesn’t seem to decline as fraudsters forge new, untraceable ways to commit their deeds. This becomes a pain for real customers who must go through needless due diligence, especially when they are in emergencies. So, it becomes indispensable to create a solution that ticks all - accurate fraud detection, simplification of the detection process, and faster outcomes. We have helped our client with all the above by designing an advanced ML-based fraud detection solution.