Running Up the Flag:

While claim adjusters quietly are going about their business, insurers are using predictive modeling schemes and rules-based technology behind the scenes to search for red flags that warn carriers when a claim could be taking a bad turn toward fraud.

Whether it is a softening market or the change in mindset that has developed in a post-9/11 world, fighting fraud still is a major focus for insurance carriers. To do it effectively, insurance carriers have continued to rely on technology tools, though not as much on software specific to fraud. For instance, Erie Insurance Group was an early entrant into the world of predictive modeling for insurance, and Dave Rioux, vice president and manager of corporate security and investigative services, has seen its use grow quite rapidly over the years, largely because of the success insurers have enjoyed with it.

Carriers are looking for software that is more universal–solutions that will help with claim processing and have some fraud benefits, indicates Michael Lucarini, global claims lead for the consulting practice at Accenture. “Making the case for a significant investment in fraud only is challenging,” he says.

One reason for the challenge is the nebulous nature of fraud. “You can't measure what you don't know,” continues Lucarini. “You can make an investment [in fraud technology], but you don't know whether you are improving your operation at the end of the day.”

Software that can impact the claim process, such as in the assignment of claims, along with addressing fraud, is where predictive modeling comes in. “What may look like a small claim today could end up as a soft-tissue claim,” Lucarini says. “You can assign [the claim] to the right person. You also can leverage the technology to determine which cases typically escalate into fraud-type scenarios. [The software] has more use than just fraud specifically.”

AutoOne Insurance, a member of the OneBeacon Group, has seen a number of benefits from its use of technology to fight insurance fraud, such as the ability to: identify suspect claims early in the life cycle of claims, resist payments on what are viewed as nonmeritorious claims, and coordinate data from a number of different sources that typically link to predictive modeling platforms. In addition, there is increased awareness and collaboration between the SIU personnel and claim personnel. “It's basically an all-in type of approach where we use a number of different resources, data points, and information to provide us with a more specific view of any particular claims or a grouping of claims,” says Harvey Aloni, assistant vice president of claim services for AutoOne.

More insurance carriers are featuring predictive modeling in their fraud-fighting efforts, believes Tony Rathburn, senior consultant for the consulting group The Modeling Agency, but it is hardly new. He points out he worked with an auto insurer in the late 1980s that was using neural networks and predictive modeling.

Probable and Predictable

When Rathburn relates predictive modeling to fraud detection, he refers to it as asset allocation. “You've only got so many people and man-hours to look at various types of transactions,” he says. “What you are looking for are the nature of the transactions, their structure, and how they are put together. Those transactions that have a higher than normal probability of something being wrong are the kinds of transactions where you want to devote your man-hours.” He explains carriers need to look for things out of the ordinary or from the past that have been investigated and have a higher propensity for fraud.

To perform predictive modeling, carriers start with their database. Models are based on historical data. “In the past, you may have investigated these transactions, and they had certain characteristics,” says Rathburn. “You have a large library of these kinds of transactions, and at that point, you start looking at other characteristics available in various fields.”

Using a variety of mathematical techniques, carriers build mathematical models with various weights on the fields, explains Rathburn. Under some conditions, there is a higher propensity for fraud, and when other conditions exist, it's likely to be a normal transaction.

Developing the models typically is done at the departmental level by someone with a fairly significant mathematical background, according to Rathburn. Once the models are developed, they are imbedded in software, and the data is run through the model to create a score, which he describes as something akin to a credit score.

Rules-Based Technology

For true predictive modeling, data models are built from the data itself so the data speaks for itself, says Rioux. Then there are systems that are purely rules based, and carriers determine the thresholds and the parameters of what they are trying to do to fight fraud. For example, if a carrier wants to look at every claim that happened within 30 days of policy inception or cancellation for possible fraud, a rule could be written to do that. “You will be getting 100 percent accuracy because it will be a simple look at dates, but does that mean there is fraud there?” asks Rioux. “We looked at the industry red flags, and what we found was when you finally get down to the red flags you actually can measure with data, [the cases] dwindle down to just a handful.”

Despite technology's help, Rioux still believes the best industry red flags are found in gut feelings and instincts. For example, if the insured avoids the U.S. mail system or understands insurance terminology better than an experienced adjuster, those should be red flags for an adjuster. “You can't measure those with data,” he says. “You have to teach those skills to the claim handler. What happens is carriers build a partial rules-based system that will flag a group of claims. Then the other red flags or the gut instincts have to kick in.”

With predictive modeling, investigators don't go in with a preconceived notion. They take hundreds of thousands of claims and analyze them. The SIU will take some known, proven fraud, look at that data, and try to find a pattern across all the other claims. The models are built from that data and then go through the algorithms.

“We have sort of a hybrid system here because we have pure predictive analytics in place with the models rebuilt on an annual basis because fraud changes,” says Rioux. “We have built in some hard rules for other things to consider. If adjusters get a claim that was flagged, and if there are any industry red flags that could be measured by data, we display those as something to consider so they have a little more information to work on. It is pure predictive modeling, but we don't ignore all the industry rules, and we continue to train and educate [people] on those gut-feeling rules you can't measure.”

In addition to rules-based technology, Lucarini notes event-based processing is being adopted by carriers because of the need to monitor a claim's life cycle when it crosses the fraud threshold, not just at the beginning or at the end. “Sometimes, at the end, it's too late [to catch the perpetrators], and at the beginning, it's not clear [if fraud has been committed],” he says. “Having event-based processing recognizes certain things.”

For example, Lucarini explains, potential scammers may contact a carrier with a medical claim but don't reveal the identity of the doctor. They know who the doctor is but won't tell the carrier for fear the claim person will draw a link. “If I run the fraud check, I won't know whether I have a connection if [the processing] is run at set intervals,” he says. “With event-based processing, you want it to run any time a new party is introduced or a relevant piece of information is entered or changes–a Social Security number, a PIN–all the elements that are needed to fill in the ring and make the connection.”

A third focus for carriers is what Lucarini calls the alert capability. “If I'm aware of certain people, addresses, or PIN numbers, I want to have an alert system running in the background that notifies me any time I get a claim with one of those elements,” he says. “Every time that element comes up–no matter what role it is in–an alert goes off, and you do an investigation.”

Front Line

Predictive modeling has served as an excellent technology tool for the front-line claim handler, reports Rioux, adding the typical claim handler is “overwhelmed” with work. Adjusters are pushed to close and pay claims as quickly as possible with regard to customer service and other expenses associated with keeping a claim open.

“You can train [a claim handler] to recognize fraud, but having an automated process behind the scenes screening claims consistently across the organization, flagging the claims, and asking the adjuster to take a second look is the direction to go,” says Rioux. “[Predictive modeling] provides consistency throughout and a greater opportunity to identify things that might slip through the cracks in a normal business process.”

Insurers spend a great deal of time training adjusters in the identification of fraud, but Rioux admits there are a number of reasons some adjusters might be in a better position to identify questionable claims while other adjusters are not doing as well. “It ranges from the volume of work they handle to the type of claims they are handling to whether they are out doing claim service in person or on the telephone,” he says. “Do they have the grasp for the fraud training so the technology levels the playing field?”

Modeling screens the claims consistently and raises to the top those claims that have a greater propensity for fraud, allowing an adjuster to focus more on those. “Not that they can ignore all their other claims, but they can focus attention on those that score high,” says Rioux.

Two Levels

Investigating claims is done at two levels with most insurers, according to Rioux, the adjuster level and the SIU level. “The adjuster level will identify a claim as questionable and require a higher level of investigation for potential fraud, and that will be recommended to the SIU, which will take it from there,” he says. A recent development has seen some companies use a different triage level where a potentially questionable claim goes to another triage adjuster level or group for further scrutiny. “There are companies looking at that, but the vast majority have two levels–the adjuster and the special investigations unit,” comments Rioux.

It is the responsibility of those two groups to identify various analytics within the core book of business to determine trends, indicates Aloni. “It's a cross-functional process,” he says. “We've created the foundation to drive better claim outcomes as a result of our ability to extract data from our own claim history with our internal proprietary software to analyze trends. This is done by various specialized personnel within each sector of the business.”

The companies that are getting fraud detection right are those that offer incentives to claim personnel, not just the fraud guys at the end, suggests Lucarini. “If you decide you want to put an emphasis on fighting fraud in your claim organization, your number of referrals is going to go up significantly, but your hit ratio will go down, so that's not a good thing,” he says. “But if you are giving incentives to people on their hit rate on the referrals, you will see some success.”

If a carrier is successful in stopping fraud, Lucarini believes more than only the fraud people should be recognized for their success. “You are seeing the most success where carriers are getting the incentive structure right, but it has to come back to both, not just the SIU unit or the auditors,” he says. “It has to start upfront with the adjuster, and the technology needs to help both the adjusters and the SIU units.”

SIU Tools

The SIU teams have a host of other tools to work with beyond predictive modeling, states Rioux. One is a data-mining tool to conduct link analysis. “It allows us to take the claim, go in, look at it in relationship to the history of all our claims, and then go to the [ISO] Claims Search system to see whether there are any relationships between this claim and others out there in the industry,” he says.

The SIU is looking to see whether connections are isolated incidents or there is some type of ring activity associated with a larger picture. “We have case intelligence analysts at the SIU level to do analytical work,” says Rioux.

The National Insurance Crime Bureau provides insurers information through alerts, as well. Erie processes those alerts against its claim information, always looking at the bigger picture. “The more opportunistic-type fraud goes off to investigators, and they will do a normal investigation into that claim,” says Rioux. Insurers are supplementing their investigations with public record information databases, using the link analysis tool and other techniques to expedite the investigation and collate the information.

Having the data compiled by ISO in its ClaimDirector fraud detection system is an asset, points out Aloni. “The system we utilize links directly to the ISO database function so we are able to combine resources from a number of different points to help us make ultimate determinations as to the identification of claims with suspicious loss indicators,” he says.

Each independent carrier has its own historical data on claim frequencies and practices, and Aloni explains that is typically where a carrier would begin to look. ClaimSearch simply is a complement to that. “It's tied to the predictive modeling piece, but by no means is it an assurance any one particular claim is or is not fraudulent,” he says.

The reporting process pops out to the AutoOne claim adjusters with specific thresholds built in within various systems. “Depending on a threshold for a particular business sector or geographic region, we may find different types of activities,” notes Aloni. “The threshold is built to necessitate a more proactive approach depending on the type of claim environment we are dealing with.”

AutoOne's systems have the ability to preserve data related to known, specific claims that previously have been investigated, adds Aloni. “If you have claims in which there are known individuals or entities who have been prosecuted by law enforcement, we preserve that information within the ClaimDirector system. So, if any future claims arise based on similar characteristics, that could create an automatic referral to SIU,” he says. “This allows us to identify [fraudulent] claims as early as possible.”

Define Success

In judging the success of a carrier's efforts to combat fraud, different measurements are used, states Rathburn. “There isn't a known amount of fraud out there, so it's hard to tell what percentage of it insurers are catching,” he says.

Nevertheless, Rathburn feels the industry is seeing a significant increase in the amount of fraud that is being detected, based on dollar values of what insurers are recouping. “Different companies have different metrics for it,” he says. “Most of the models I look at in the fraud area are dollars returned per man-hour. A number of companies are looking at total dollars saved. I've also seen companies look at the number of cases. It depends what their priority is. Some are not that concerned with the dollar numbers; they just don't want [fraud] to get out of hand. They want to go after as many inappropriate transactions as possible to try to make sure they are capturing everything, which is going to diminish the dollar per hour of investigative time.”

However, the typical ways of measuring effectiveness against fraud, such as referring more cases to the SIU, are not a good indicator. “You can look at whether you won more cases or successfully denied some cases,” says Lucarini. “That helps, but you still never know what that was a percentage of. Did you kill five percent of it? Ten percent? All the measurements are based on cases that have been found.”

There are several metrics used by Erie to define success, but Rioux reports the carrier basically is looking for missed opportunities in claims that were either overlooked or not identified early on as having fraud potential. “Predictive modeling and other technology we've used have allowed us to identify [fraudulent claims] early on and proven time and time again there were missed opportunities,” he says. “When we use predictive modeling and score claims, particularly claims that were not referred or identified early on, we find there was an opportunity the claim should have been investigated further because it had the characteristics of a fraudulent claim. We've been able to go back and document that on a number of occasions.”

The SIU units always will have to defend their existence, according to Lucarini, usually based on how much the unit costs the insurer and how much it saved the carrier in stopping fraudulent payments. “In the end, that's the age-old measure we come back to,” he says, adding fraud detection is cyclical with the market. “If [carriers are] fat and happy and having a good year and they are able to raise rates, there's less emphasis on fraud,” says Lucarini. “When you are in a tight market where you can't raise rates and the company needs to save organically, there is more emphasis on fraud. It's cyclical with the softening of the market.”

Best of Both Worlds

For years, carriers used hard rules because they could leverage the red flags that had been around for 100 years. However, over time, while companies have found they were getting 100 percent accuracy with tracking claims with their rules-based systems, they also were getting plenty of false positives. “These claims met all the conditions but clearly were not fraud by any stretch of the imagination,” says Rioux. “When you start looking at the industry red flags, fraud continues to change, and what is neat about predictive modeling is it's more dynamic and intuitive. It gets more intelligent each year you rebuild the models.”

Models are based on past experiences, so if claim adjusters get a bunch of referrals throughout the year based on their predictive models and they've had success in proving fraud, carriers are going to rebuild their models based on some of those successes so the model gets a little more intuitive.

On the flip side, if there are claims that probably don't deserve an SIU investigation or aren't suspicious, a carrier can rebuild models based on that information so the carrier would be getting fewer referrals to SIU. “With predictive modeling, we are getting referrals faster than we've ever gotten them in the past,” says Rioux. “That reduces the claim life cycle. We've also found the ones flagged by the system and investigated have demonstrated the greatest propensity for fraud. We've gotten more from that population than we've gotten from the normal referral process. We've also been able to identify more fraud than we previously had with manual processes. It continues to demonstrate the models are doing what they should be doing, and that's important.”

Decision Support

“What we do really is decision support,” says Rathburn of predictive modeling. “We are playing a game, and you keep score with a bank account. All we are doing is developing a mathematical model that lets [insurers] make better decisions so they play the game better and improve their score.”

There still is a place for the tools that focus specifically on fraud, Lucarini believes, but they need to be integrated into the business process correctly rather than serving as stand-alone tools. “If you can integrate those tools with event-based capabilities or with rules-based capabilities, you are maximizing the value [the software] can bring to the table,” he remarks. “I tell people the last thing I want to do is be a fraud software salesman. I can go in and tell people it will save them $10 million this year, and it's probably true, but you just can't prove it.”