Businesses today understand the importance of big data and analytics and have started embracing it to adapt to the changing customer expectation. And along with it this, the pressure to reduce costs and realize operational efficiencies is on an all-time high.
Fraud appears to be an ideal case for applying advanced analytical techniques - whether it is on the history of fraudulent claims, identifying claim types, demographic patterns, policies, claim patterns and violation - the Industry has a model that can score an incoming claim and minimize fraud.
Developing an analytical model basis past fraud and operationalizing the developed model to effectively minimize future fraud is anything but straight forward and would typically lead to:
- High false positive in terms of fraud prediction and changing fraudulent pattern
- High false positive resulting in triaging overheads
This should not come as a surprise. Number of fraud cases as a percentage of the total claims should typically be small, say less than 10%. Numbers may be even small in absolute numbers, making it very difficult to develop any scoring model. But analytics is not all about building complex statistical models. We could develop rules based on the domain knowledge, past experience, local knowledge and patterns extracted from historical claims data and then use the rules to identify a critical fraud case that needs to be referred to the Special Investigation Unit SIU. Some examples of the rules are:
- Claim within 2 weeks of policy inception and claim for fire or theft
- Claim within 2 weeks of policy coverage change
- Witness and Third Party have same or similar address
- Fault claims with multiple third party from same area (postal code hot spot)
All claims are scored against the set of rules and claims with a score beyond a threshold are flagged for review.
The rules engine drives referrals and the outcome helps to reduced false positives and increase capture of fraudulent claims. Predictive analytics/models can then be deployed to identify rules producing false positives and then reduce the false positives by amending the rules engine to ignore false positive combinations of rules. Combining predictive model with rules results in better rules management and less false positives in captured fraudulent claims.
Typical expectation would see fraud ratio in the region of 8% - 12% of new claims being referred. Clearly, the success of any model is dependent on the quality and availability of data to base it on and the flexibility of the rules base to ensure that the model can be maintained as the nature of fraud changes within a book of business.