Predictive Fraud Fighting

Wes Gill, Executive Lead and Head, Governance, Enterprise Risk and Compliance; SAS Institute (Canada) Inc.
James D. Ruotolo, Principal, Insurance Fraud; SAS

Earlier this year, the Insurance Bureau of Canada (IBC) issued a release entitled “They cheat, you pay – Insurance fraud on the rise in Ontario.” In it, IBC discussed the complexity of today’s insurance scams. It also indicated it focuses its efforts on investigating organized crime rings involved in, among other things, fraudulent auto accident claims.

Fraud is certainly a prevalent problem. Thirty-nine per cent of senior executives surveyed last year by SAS/Leger said customers had attempted to defraud their organization. But once organized crime rings get involved, the situation becomes far more complex and costly. Crime rings are active, persistent, methodical and quickly adjust their efforts to improve their odds of beating the system.

Historically, insurers have relied on adjusters to help identify problematic claims. Adjusters are uniquely talented at identifying claims anomalies, and this manual approach will always be an important part of any insurer anti-fraud program. However, it becomes increasingly difficult to identify ‘red flag’ situations manually when, on their own, individual claims might not show any signs of being problematic. Sometimes alarm bells will be raised only when seemingly disparate cases are connected. In addition, the sheer volume of information now available to adjusters makes detailed analysis very cumbersome and time-consuming. Insurance fraud rings work to capitalize on this situation. To fly under the radar, they submit claims that do not seem out of the ordinary.

Fighting Fraud

In light of the above, many companies are looking to technology to help them identify suspicious claims. “Suspicious” claims not only involve what is being claimed, but also who is making the claim and when and how the claim is made. The goal is to identify patterns and assess whether individual cases show signs of fraud based on their commonalities with other seemingly unrelated cases. The subtle similarities might escape even the most seasoned adjuster.

These connections are often the Achilles heel of crime-ring-based fraud: there are only so many degrees of separation existing between all the participants. This is where monitoring people’s social networks can be helpful. For example, certain combinations of information may indicate a fraud attempt and require further investigation. Perhaps individuals with similar-sounding names frequent certain repair garages; maybe this ties in with a small network of household addresses combined with a cell phone number that pops up several times – once as that of the driver in an accident and the next time, under a slightly different name, as that of a witness. Social network analysis helps uncover these previously unseen links.

Integrating social network analysis tools into a broader fraud framework is key to an organization’s ability to guide adjusters and optimize efforts to detect fraudulent claims. By using a framework with a comprehensive fraud scoring engine that incorporates a combination of different analytical techniques – automated business rules, database searches, anomaly and exception reporting, predictive modeling, text mining and network link analysis – adjusters are able to determine the likelihood a claim is fraudulent and prioritize their efforts accordingly.

Even if a corporation shows a well-established desire to move forward with predictive analytics, the quest to find the best solution may present hurdles thwarting companies from moving forward with their plans.

Excuses, Excuses

Here are the Top 5 data excuses for not using predictive analytics for insurance fraud detection – and why they’re wrong.

Excuse #1: We don’t have enough data

Standard approaches to predictive modeling for insurance fraud detection involve analysis of an existing set of known suspicious claims. From this data set, it is hoped predictive indicators may be found that could be used to identify similar claims in the future. This technique is very powerful, but it relies on a large set of known suspicious claims on which to build a model and train people on how it works. If a company has limited known suspicious claim history information, it often believes that it cannot proceed with a technology-assisted fraud detection program.

Reality: A number of statistical approaches can be used to build a solid predictive analytic solution, even if very few suspicious claims have been identified in the past. For example, a hybrid solution combining business rules, anomaly detection and social network analysis can identify suspicious claims even if no suspicious claim history is provided.

Excuse#2: We don’t have good data

Overworked adjusters and claim processors have a tendency to find the path of least resistance in order to meet their objectives. Ever discover a claimant with the Social Insurance Number of 999-999-999 or an address of “Unknown” in your claim system? Data quality issues are a reality for any large organization. Many analysts and investigators have been frustrated by poor-quality data in transactional systems. They often feel data quality problems may prevent them from being able to implement a predictive analytics solution effectively.

Reality: The old adage “garbage in-garbage out” certainly holds true for many things, but data quality issues do not preclude a successful technology implementation. Simple tools like basic business rule engines may be less effective in dealing with data quality problems, but a robust insurance fraud detection solution must incorporate data preparation steps that carefully cleanse the data to remove problems. However, be careful not to clean too deeply. Improper data cleansing techniques can actually harm the data set, by erroneously categorizing anomalies due to fraud as data quality errors to be removed.

Excuse#3: Our data is too fragmented

Information silos are prevalent in the insurance industry. Business units are beginning to see the value of sharing data across the enterprise, but many organizations house and manage their own data. Given the fact that most companies use a patchwork of transactional systems for ratings, customer service, policy administration, claims administration, payment processing and human resources, it’s no wonder that their data is fragmented. With all of this information located in different places, fraud detection projects are often shelved because they are perceived as too complex.

Reality: It is not necessary to revise the entire corporate information technology infrastructure to build a fraud detection solution. Enterprise solution vendors can leverage data integration tools to incorporate key data elements from various internal systems. By combining information from these disparate information sources, new insights and fraud detection capabilities are immediately possible. For example, one insurer has been able to cross-reference data from multiple claim systems for different lines of business. In another example, one commercial policyholder found more employees were involved in claims on the auto policy than were covered on the workers comp policy, suggesting a possible premium fraud scenario.

Excuse#4: It’s all in the notes

Some studies suggest that upwards of 80% of insurer data is unstructured text. Any SIU investigator will tell you that the most valuable information about a claim is not in the discrete structured data fields – it’s in the notes. It’s impractical to have a unique field for every piece of useful information; as a result, the claim notes become a rich information source. But text fields are not generally used for reporting purposes, and therefore are not often available in data warehouses. They are therefore not cons
idered a viable data source for a predictive model.

Reality: Text analytics can be one of the most powerful components of a hybrid fraud detection approach. For the same reason, any seasoned investigator will want to read the claim notes. A predictive model should make use of unstructured text data. Entity and variable extraction are fairly straightforward using advanced text mining tools. In some solutions, up to half of the data elements used in a predictive analytics fraud detection solution comes from unstructured data sources.

Excuse #5: We can’t handle any more cases

SIUs often have limited budgets and have to maximize their scarce resources. Most of them already have more work than they can handle. When asked about a predictive analytics solution to identify suspicious claims for further investigation, some organizations protest. “No thanks, we are already swamped,” they say.

Reality: Technology can help organizations identify more cases to investigate, but that’s not the only benefit. A critical, but often overlooked benefit of a fraud technology solution is the ability to prioritize work more effectively. Most organizations operate on a first-come, first-served basis; they simply work on the cases as they come in. Business rules, reporting tools and case management systems can help SIU leaders better manage their scarce investigative resources. Even if it’s investigating the same number of cases, an SIU can dramatically improve productivity and impact rates by effectively prioritizing caseloads.

The Bottom Line

You are likely to encounter naysayers in any organization who will offer the above excuses, but don’t be discouraged. These challenges are easily overcome with a robust, technology-assisted insurance fraud detection solution.

The bottom line is this: given the increase in fraudulent claims, it is imperative for insurance companies to leverage technology as a key enabler to combat crime-ring-based fraud. Organized fraud is, by its very nature, active, methodical and extremely agile. By leveraging both structured and unstructured data, firms will be able to determine the likelihood a claim is fraudulent, prioritize their efforts and reduce their claims expenses significantly.