Fraud Detection for a leading Personal Lines Insurance Company

Overview
Fraud analytics has become a pivotal aspect in the Insurance industry. In the past, fraud detection was majorly heuristics driven. With the advancement in fraud analytics, a more structured approach to fraud detection is being employed in multiple areas, such as underwriting, policy renewals, regulatory compliance, claims review etc.
False claims, intentional falsification of data during the underwriting process, and fraud perforated by crime rings cost billions of dollars in losses to insurance companies each year.
Existing fraud management practices at a leading personal lines insurance company had room for improvement given the significant need for manual intervention to identify fraud and a high rate of false alarms in claims being flagged.
Mu Sigma helped the insurer improve fraud detection methodology, resulting in ~$30M potential savings.
The Problem
The existing approach for fraud management employed by the client was largely business rules driven. The existing fraud detection model was ineff¬ective in identifying suspicious patterns and inflated claims through examining historical data.
Further, the approach relied on manual intervention to understand patterns among claims having very similar or abnormal descriptions. Hence, numerous actual fraudulent cases were not being caught and a significant amount of the investigators’ time was spent chasing false alarms.
Mu Sigma Approach
Mu Sigma had been supporting the claims team for a couple of years to improve the efficiency and effectiveness of the claims process. Being familiar with the claims life cycle and data, we followed the following approach:
Text Mining
One of our hypotheses was that text entries captured during various stages of the process could reveal scripted patterns indicative of suspicious fraudulent activity and thus help in fraud detection.
Since these were not easily visible during the manual review, a lot of these cases were not being flagged. Mu Sigma built a text mining algorithm and generated a fraud propensity score that combined the business rule classification with the authenticity of textual entries.
Testing with the investigation unit in the field revealed that the number of fraud cases being caught improved by 5% over a control group. Some of the suspicious text patterns found through the model indicated collaboration between various entities involved in the claims process, which the field unit wanted to investigate further.
Social Network Analysis
Borrowing the concept of influence mapping used in social media marketing, Mu Sigma performed a social network analysis to identify suspicious relationships between various parties involved in a claim (claimant to employers, claimant to medical providers, claimant to legal representative firms etc.).
This element was added to the propensity score. While this further improved the number of cases being flagged, the number of false positives was still high and was a cause of concern for investigators.
Predictive Modeling
To address this issue, we listed all the possible factors that could be indicative of suspicious activity and identified relevant data elements to be used through our problem-solving framework – aops. The identified factors were then used to build a logistic regression model that helped predict the probability of fraud given historical customer characteristics and the behavioral characteristics from known cases of fraud.
This model replaced existing business rules and helped improve the accuracy of the cases flagged. A composite fraud propensity score was then generated using all three elements– text mining, social network analysis, predictive modeling. To enable ease of consumption, a claims case tool was developed and deployed that allowed investigators to visualize and examine the fraud propensity score along with relevant factors/reasons for suspecting fraud in each case.
The Impact
The improved approach was able to capture 90% of actual fraud cases by investigating a sample set of total claims payments.
Implementation of the model in test markets showcased potential savings of over $30MM through the identification of 20% more fraudulent claims while flagging a smaller number of cases for investigation.
Click here to know about how we helped a leading financial services company reduce transaction monitoring process investigation time by 50% and highlight fraudulent entries.