The numbers are huge: an alleged $297 million insurance fraud case, with 36 defendants, including 10 licensed doctors, three attorneys, 22 medical professional corporations, and eight members and associates of a criminal organization. All charged with offenses related to the single largest alleged no-fault auto insurance fraud case ever, and the first to include alleged violations of the Racketeer Influenced and Corrupt Organizations Act (RICO), which have been filed against the eight.
According to the U.S. Attorney's office, from at least 2007, the 'No-Fault Organization' is said to have defrauded auto insurance companies by creating and operating medical clinics that provided unnecessary and excessive treatments to take advantage of the no-fault law.
While charges are still pending, the No-Fault Organization case is extraordinary because it's indicative of the increasing sophistication and pervasiveness of insurance fraud, which is on the rise. According to the latest statistics from the National Insurance Crime Bureau, the number of questionable claims in 2011 increased 9 percent to 100,201, from 91,652 in 2010, which was an 8-percent increase from the 84,845 questionable claims reported in 2009.
Insurance fraud costs the U.S. property/casualty industry an estimated $40 billion to $120 billion per year, says Stephen Applebaum, senior analyst, property/casualty insurance at Aite Group. "It's hard to measure money that leaks out when you don't know where it goes," he explains. "But in an industry that only generates $495 billion annually, it doesn't matter whether it's $40 billion or $120 billion, it's too much."
But, as fraudsters become more organized and methodical in their methods, so too are insurance companies. To detect fraud, insurers such as Allstate, CNA, Erie Insurance and others are employing a raft of technologies-from rules-based analytics included in policy administration and claims systems, to predictive analytical models that scour internal databases, free-form text and external databases, and now social media sites-to identify patterns and anomalies indicating organized and sometimes massive fraud, for review, investigation and potential prosecution.
"About 18 months ago, Allstate drew a line in the sand," says Frank Llende, senior manager, Allstate innovation and field support. In addition to filing a $29.9 million lawsuit against the No-Fault Organization's defendants, the insurer is keeping an eye out for other fraud. "This is organized crime. We were convinced that there's got to be a way to identify anomalies, trends and aberrant behavior among organized and connected individuals. We need to be able to see who they are, what their behaviors are, and confront them as quickly as we possibly can and get this information into the hands of law enforcement and send these people to jail."
Use of Tools
The industry has come a long way. In the old days, the claims process was linear. "Every phone representative would take every claim in exactly the same way and pass it along to another unit that decided which claims unit needs to be assigned to it," Applebaum says.
Claims adjusters were then left to identify red flag fraud indicators on their claim files. Claims that accumulated enough red flags then were referred for investigation. The problem was that some adjusters are better at recognizing fraud, and as a result many referrals could originate from a small number of adjusters. The company also couldn't know how much fraud was missed.
Rules-based and predictive analytics can be used to detect fraud much earlier in the insurance lifecycle. When a policy is written, for example, companies now are able to access external databases to locate undisclosed drivers, such as spouses and children who share an address; access driving records and police reports; or find whether a driver has been dropped from a previous insurer, all common opportunities for fraud.
At first notice of loss, they can help an insurer beat the clock on paying claims. "The questions go: 'Were you injured? Yes. Was anyone else injured? Yes.' And in the background, the analytics engine is working: 'We've got a bodily injury and a third-party bodily injury. We are going to rout this claim to unit 1-A, because that's the unit that is highly skilled with third-party bodily injuries,'" Applebaum explains. "What the analytics have done is put the claim in the hands of the people best trained to handle it, compressed the claims cycle by days and increased the customer satisfaction level, because the policyholder is getting treatment and satisfaction in a shorter period of time."
Rules-based systems score claims based on an accumulation of red flags from data in claims and policy admin systems, explains David Rioux, VP and manager of the corporate security department for Erie Insurance. In and of themselves, each flag may be innocuous, but as the claims file develops and more data is added to it, claims that collect enough red flags are then reviewed and potentially investigated.
"Those red flags may be, if it's a workers' comp claim, a Monday-morning injury. That could mean someone hurt themself over the weekend, dragged themself to work and had an accident in conjunction with short-term employment," Rioux explains. "Or in auto, there could be a lien holder interest, and the thing ends up stolen."
Rioux lists a string of examples that add flags to a claim. "But it's never just one thing. There will be a combination of factors that will either add or subtract from a score, but we tweak the threshold," he says. Set too low, fraud goes undetected, set too high, and the insurer ends up with too many false positives. "It's really about trying to find the sweet spot," he says.
The benefit of rules-based systems is that they can raise the effectiveness of less experienced claims adjusters. "We have some claims adjusters who have no trouble identifying questionable claims," Rioux says. "They are in tune with the facts and circumstances and can pick them out. We have others who work higher-volume desks with lower limits, or don't have the experience or tenure in the claims industry."
Next Generation
While the rules-based approach continues to produce results, predictive analytics, which incorporate data mining and statistical modeling, increasingly are finding acceptance with larger insurers.
At Erie Insurance, Rioux uses FraudFocus, from LexisNexis, which incorporates a hybrid of rules-based and predictive-analytic data modeling. In addition to a rules-based engine that looks for individual questionable claims, Rioux's group is able to detect questionable claims that have the same or similar characteristics of past, known-fraud investigations.
The company tracks fraud cases and builds predictive analytical models based on historical data: predominantly on known outcomes of questionable claims, for each of its lines of business. The model looks at fixed fields and free-form text in claims and policy management systems, such as vehicle makes, models and years, date-of-loss, time-of-loss, type of loss, initial reserve amount and other basic attributes, looking for connections, recurrences and other anomalies.
"Those models run around the clock, analyzing claims data and they do it at first notice of loss, electronically scoring and monitoring throughout the life of the claim," Rioux says. "That claim may have a predefined threshold when it will flag. And it may never flag; it could flag on the first day, or a year later. Let's say there's a third party added to the loss and that threw the model over the threshold to flag, or another vehicle or a secondary claim was introduced; different factors could change the model, but it is constantly being scored."
Founded in 1897, CNA now fights fraud with its major investigations unit, which is armed with SAS Fraud Framework, a hosted solution. "We feed all our open claim data to SAS and they host it for us," says Tim Wolfe, director of CNA's special investigative unit. "We also send SAS information from our own SIU case management system so they can see where we have referred cases for prosecution in certain states."
Through a user interface, SAS then returns a list of claims, which have not already been referred by adjusters, scored for fraud potential. CNA receives two types of scoring, those indicating a suspected individual fraud, and provider alerts, which indicate the possibility of more complex networks of fraud. The CNA team then reviews and triages the questionable claims, prioritizing them by severity, and determining whether they should be pursued and investigated.
For many insurers, text analysis is a relatively recent addition to the fraud detection toolbox. Search algorithms run against the open, unstructured data frequently included in an adjuster's notes in a claim file log, searching for certain words and phrases common to previous fraud investigations.
"It could be things like the adjuster makes a notation that there are inconsistencies in the statements they are taking, that the claimant refuses to cooperate with the investigation, or refuses to give a recorded statement," Wolfe says. "The model hits on those and uses them to help generate the fraud score."
Looking Outside
The power of the predictive analytics has been magnified by link analysis and the inclusion of external databases. Insurers often incorporate public records data in their models, such as state databases of professional licenses, disciplinary actions and sanctioned medical providers.
"We can get things such as corporate entities, so we can find out who is behind those organizations. One of the things they look for is doctors referring patients to entities in which they have a financial interest," Wolfe says. "That can be very useful. Also, civil and criminal litigation records-we find that people who commit insurance fraud have a propensity to sue and be involved in criminal activities. So that's something we look at very closely."
For auto claims, Erie looks to LexisNexis for information on the vehicles involved in the crash and the parties to the loss. "The external data we are introducing is: parties to the loss against known criminal felony records, that may help provide insight," Rioux says. They also look for questionable claimants with bankruptcies and foreclosures. "You are not going to run this against everything, but if you can automate it and incorporate it into that model, and let the model do its magic, you may give the model extra 'lift.' That's where these models are heading."
While it is becoming standard practice to search for signs of financial distress, such as liens, bankruptcies and judgment reports, there are limits to what insurers can and do access. "We follow fair credit reporting guidelines and Gramm-Leach-Bliley and the Drivers Privacy Protection Act," says John Kloch, director of Sentry Insurance's special investigations unit, and the company doesn't access credit reports, for example.
Putting It Out There
Insurers also are using and contributing to other databases such as such as Verisk Analytics' ISO ClaimSearch in order to better understand certain loss scenarios and trends for individual business lines at the national level.
"We already know that within our own claims data there are suspicious factors about this certain entity. When we incorporate the ISO data, we find out if that entity is also on the radar screen and has filed other claims with carriers like us. That will generate additional leads for us to follow," Wolfe says.
Insurers submit claims to ISO and then have the ability to match their own claims data against claims filed with other carriers. It's not unusual to find the same claimants, addresses, cars, phone numbers, doctors, medical clinics, show up with frequencies that indicate fraud more than just bad luck.
"That's where the data linking comes in," Erie's Rioux says. "You can take in thousands of claims and determine who are the ring leaders, or what data attributes have the highest number of linkages."
Another source insurers, including Allstate, turn to is the National Insurance Crime Bureau (NICB), which partners with insurers and law enforcement agencies to aggregate data about questionable claims and issue alerts about fraud activity, and facilitate the identification, detection and prosecution of insurance criminals nationwide.
"We bring those alerts into our systems, and we take our data and the hypotheses the NICB has created: 'providers X, Y, and Z, with this particular body shop;' and find our exposure," Allstate's Llende explains. "There are a couple potential situations. One is that we are in fact infected with the same fraud, so we will take that case on. Or we may validate that we have that case in place. Or, we may not have that problem at Allstate, but now we know about it, so we can build that into our rules and models, so we are prepared to take that case on."
Seeing is Believing
In order to understand the enormous quantities of data and grasp the connections between people, places and things that come out of these fraud detection systems, insurers use data visualization tools to present their results to analysts, investigators, law enforcement and juries.
"If you can see the picture, you can see the fraud," says Brian Smidt, VP of data analytics at NICB, adding that they use ISO's NetMap Analytics for visual link analysis and IBM's i2 for discovering networks, patterns and trends in both structured and unstructured data. "It's much better than dealing with massive spreadsheets. It's a game-changer because you have a way of looking for fraud that you didn't 10 years ago."
A link analysis chart looks something like a spider's web, charting the relationships between all of the parties to a claim, and potentially several claims. At CNA, the link analysis is reviewed and triaged by a team of data analysts within the major investigations unit.
Using the SAS Web interface, analysts drill down into the case, where they can see a list of the red flags associated with the case and why it has scored high for fraud potential based on business rules, predictive analysis and anomaly detection at both the individual case and fraud network levels. With the SAS Social Network Analysis, the analyst can see when people, addresses, vehicles and other claim details recur across multiple claims.
"It's important to note that the tool does not do the investigation for you. You still need an investigative specialist to do the field work or drill down to the next level, to actually uncover the evidence of impropriety, if there is any," Wolfe says.
Creating easily interpreted, actionable information is especially important given the enormous amount of data, from multiple sources, that goes into a fraud investigation. "It's really about putting the data together so that it makes sense to an investigator who has a very large pending [folder]," Allstate's Llende says. "It's triage, it's prioritization, and it's allocation of resources, using data to inform those decisions."
At Allstate, analysts use Tableau to explain and present the information to adjustors and investigators, and to find out-of-pattern trends between and within geographies, for example. "The ones that are going to win are those who can do change management down at the desk level," explains John Reardon, director claims centralized services at Allstate. It also requires working in a team environment, bringing many data points together so staff can look at it, understand and interpret it, and engage with law enforcement, he adds.
"The claims professional is really developing. If you think about where we were 10 or 15 years ago, we never had individuals who were Ph.D. statisticians, predictive modelers and data experts, and people who could deal with big data," Llende says. "Now we are bringing all those types of people into the process."
In addition to the personnel, the data creates its own series of challenges. The first is capturing the right data points, which if they exist, may exist outside of the claims system. "The second point is the legacy systems," Llende says. "The companies that are going to win are those that can not only work in their contemporary claims systems, but also go out and grab historical legacy data, make sense of it and marry it up to disparate data and platforms."
Disparate data and legacy systems can present a serious challenge. One of the perils of legacy systems is that the accumulated components were written separately, and at different times, sometimes spanning decades, each with unique attributes. "The person who wrote them was probably a genius in making them all talk to each other, but they are still all different components. And if you can get to a one-component platform that is Web-based, that all uses the same coding language, database and that, it's much easier," Llende says, adding that as companies build out legacy systems, determining how to access the data should be part of the architecture.
In many cases, the data itself may not be ready for any sort of analysis. "Nobody ever thought we'd want to use that [claims and policy admin] data for analytics," Erie's Rioux says. "We have these old legacy systems that have 'first name last name' in one field. So you have to parse it out. Or the whole address is on one line. You struggle with this dirty data to clean it up and make it usable."
To overcome those sorts of challenges, Allstate's Reardon says the company went through a major change, replacing a significant amount of the legacy-based system with Accenture Claim Components, called NextGen internally, which is a Web-enabled technology that gave them greater access to the claims data. "It's accessible, it's complete, it's consolidated, it integrates to the point that we can move it into these SAS tools and other things we have developed."
The future of fraud detection is likely to include automated searches of social media, NICB's Smidt says. "People who commit murder don't usually put it on their Facebook page; people who commit insurance fraud don't think twice about it. It's ridiculous but true," he laughs, adding that such pictures frequently include metadata that includes when and where a picture was taken.
In addition to detecting fraud, the future will also entail using these same tools increasingly to improve customer care by speeding claims payment for innocuous claims, increase opportunities for subrogation and reduce premiums by reducing fraud.
"Our ultimate goal is to put these people out of business, but we will settle for pushing them to other companies," Llende says.
