≡ Menu

Detect, prevent, manage claims fraud


Insurance claims fraud is a huge deal, with most estimates seeing perhaps 10% of claims made being fraudulent – worth at least $30Bn/year. As a result the value of small improvement is huge. Of course, from my perspective, fraud is one of the classic uses of decision management and predictive analytics. Obviously the current economic situation tends to drive up fraud, especially given the rate of job losses and the likely lag before the employment rate improves. Add in the fact that nearly 25% of home owners have negative equity and the very high rates of delinquency in subprime mortgages and the rate of fraud is likely to be high for some years to come. Some categories of claims have risen by 25-30% since last year. Not everyone thinks about insurance fraud the same way as they do about crime – they think that it is OK or marginal to add to or otherwise exaggerate claims, for instance.There are also many kinds of fraudsters – those who don’t want to commit fraud but are encouraged to, those seeking revenge on the insurance company because they have been paying for a long time, those who like to game the system and ultimately organized crime rings both internal and external.

At the end of the day about 5-10% of insurance premiums are driven by fraud. Given the low rate of fraud detection in many companies, getting serious about fraud can really drive big improvements in loss ratios and profitability. Improving the loss ratio and hence the overall combined ratio is critical as it drives profitability when investments are doing poorly. Today far too few are using analytics, especially too few using analytics integrated into operational systems.

The typical claims processing model begins with First Notice of Lost and then an assignment step, a decision service in my terms, that routes to automated settlement, investigation or evaluation. Analytics can really help with this. SAS, like me, sees business rules and predictive analytics being used together as part of this decision. All of this is wrapped up in SOA so it is easy to link to legacy claims systems or business process management systems. Once you start investigating or evaluating then there are other analytics, like social network analysis to find fraud rings, to help with this kind of decision support. Reserving, subrogation and others all also have potential analytic opportunities.

SAS has found that the critical need is for agility and responsiveness because fraud is such a dynamic and ever-changing problem. Fraudsters are increasingly organized crime rings and they keep looking for new ways to defraud companies so companies must be able to respond equally quickly. You need also to balance rigor and responsiveness, customer service and fraud detection. Analytically prioritizing claims, then, becomes critical. If you can detect more fraud without doing more investigations, without referring more claims to the investigation unit, then you get a boost to the bottom line while improving customer service.

Moving on to the technology, rules are a baseline – essential to processing claims and finding fraud – but they can become hard to optimize and trade-off and they can be overly fixed. Unsupervised learning, things like neural nets, compare claims with peer groups to find anomalies and outliers. Taking known-fraud cases too, and building predictive models from them really helps too. Now you can flag those that look like known fraud, those that are outliers and those that fail basic rules checking to find fraud. Social network analysis can also be used as a driver of this automated claims processing – the engine is more likely to flag claims from someone who appears to be in a fraud group for instance.

Once you find fraud you can add things like social network analysis to track down the rings. Obviously this helps with organized rings but those that know someone who has filed a fraudulent claim are more likely to file a fraudulent claim themselves so it helps with implicit rings too. Finally all this feeds case management and alerting. Lots of tools for visualizing why a claim has been referred for investigation, the networks of people behind the claims etc. It is important to have good linkage of decision support and decision management, presenting the same analytic insight differently for automation and manual reviews.