Data Mining Improves Fraud Mitigation Efforts
Though not there yet, Los Angeles-based Farmers Insurance, a subsidiary of Zurich Financial Services ($67.2 billion in revenue), is moving rapidly toward adding text-mining capabilities and predictive modeling to the front end of its claims and fraud-fighting platforms, says Doug Ashbridge, director of special investigations at Farmers. Text analytics solutions extract factual information from unstructured text found in documents such as police reports, medical records and adjuster notes to establish patterns and identify trends.
 |  |
Doug Ashbridge, CIO, Farmers Insurance | |
|
The Farmers claims system has hundreds of structured data points, but Ashbridge still believes that a text analytics solution's ability to mine unstructured text could be beneficial to the organization's fraud-mitigation efforts. "By mining the text, you come up with more data points that were missed," he explains. "It also allows you to look for scenarios that are verbalized, such as the way in which an accident took place. You can't grab something like that out of data points."
Michelle de Haaff, vice president of marketing and products at Attensity (Palo Alto, Calif.), a provider of text analytics solutions, including one specifically designed for workers' compensation claims, says that most insurers do not have processes in place to look at the unstructured data that represent a large portion of the total data housed within a claims file. "In general, an insurer may have excellent processes in place regarding fraud, but they're missing a big, important piece that makes a huge difference in their ability to identify fraud on a mass level," she contends.
While Farmers may soon implement data-mining technologies such as text analytics, it already has an active and effective fraud-mitigation operation, according to the carrier's Ashbridge. In late July, for instance, the carrier won a fraud lawsuit against Hollywood Auto Collision. The case was the first tried under a 1993 California statute designed to augment law enforcement efforts to prosecute defrauders, according to a Farmers press release.
Zero-Tolerance Policy
Farmers enforces a zero-tolerance policy regarding fraud, according to Ashbridge, who has been associated with the company's SIU for 25 years. While acknowledging the organization's aggressive approach to fraud investigation and the diligence of its well-trained SIU and claims adjuster workforce, Ashbridge credits technology with helping to improve fraud mitigation over the years.
"We have a very robust back-end process where we're able to pull in large amounts of data from our databases, massage that data and show trends and patterns -- areas where we should start to look for fraud," Ashbridge says. Currently, he relates, the company uses ISO's (Jersey City, N.J.) link analysis technology NetMap and i2 investigative analysis software to quickly identify trends and patterns within its own past claims data and information drawn from outside databases, such as the consolidated All-Claims database cooperatively maintained by ISO and the National Insurance Crime Bureau.
McLean, Va.-based i2, which is owned by ChoicePoint (Alpharetta, Ga.), allows Farmers to graphically visualize claims patterns, Ashbridge adds. "NetMap pulls all the information together and i2 allows you to display it in a way that makes sense," he says.
Accuracy First, Detection Second
But making sense of the data is harder than it sounds. Analyzing data to detect trends and patterns -- as opposed to simply using old fraudulent claims to directly detect new fraud cases -- is only now an emerging industry best practice. "There is a certain degree of fallacy in using your past fraudulent claims to predict future claims," Accenture's Costonis says. "The key thing you should be trying to recognize is a pattern and how it evolves over time as opposed to trying to predict the specific fraud indicators."
Columbus, Ohio-based Nationwide ($22.3 billion in 2006 consolidated total revenue) employs a similar philosophy and utilizes several different technologies to defend against fraud. The insurer uses ISO NetMap for links analysis capabilities and, in 2005, implemented a solution from Chicago-based Magnify (another ChoicePoint company) for predictive modeling, according to Donnie Kearns, SIU director, Nationwide.
Kearns says Nationwide leverages predictive modeling, in part, to find trends in past fraudulent claims. "Leveraging predictive modeling improves the accuracy of fraud detection. In short, it narrows our focus on fewer claims," he explains. "Claims professionals can't really appreciate patterns because they are handling so many claims, and each is treated as an individual transaction. They don't get a chance to step away from the forest to see the trees, so to speak."
Kearns and others caution, however, that using only predictive analytics to detect fraud could be ineffective. "As powerful as modeling is, you shouldn't let that work alone," he stresses, adding that fraud mitigation is more than just learning from past mistakes, especially when it comes to predictive modeling, which uses past evidence to predict future results.
"Predictive modeling is primarily going to tell you what you know or have experienced," Kearns continues. "You shouldn't stop there but, rather, employ other tactics. Those efforts should make your model stronger over time."
And few observers doubt that current fraud predictive models have room for improvement. According to SPSS' Parker, many experts within the insurance industry believe that 12 percent to 15 percent of all claims should be referred to the SIU as potentially fraudulent. However, only a small percentage of those are ever actually identified as such.
"That means that when you go to create a predictive model based on that half of 1 percent, you're not including all those [fraudulent] claims that you failed to identify," Parker explains. As far as the predictive model is concerned, those unidentified claims could be considered valid, throwing off the accuracy of the system.
Even the SPSS solution suite itself, known as PredictiveClaims, does not rely entirely on predictive modeling, Parker notes. He says the solution first scores claims based on a business rules engine and then "optimizes" those scores with a predictive model.
Tweet This