To enable more effective industry-wide collaboration, individual carriers must develop the right kind of corporate culture, create practical processes for handling suspicious claims, and implement systematic fraud detection solutions. Corporate culture is arguably the most important. Fighting fraud must be an enterprise-wide endeavor, with commitment beginning right at the top. Investigating suspicious claims cannot be a unilateral decision of the claims department because it also inevitably has implications for the legal department (in cases of disputes and suits); business areas (lost policyholders from incorrectly referred claims); and public relations (overcoming negative publicity such as blogging from dissatisfied claimants).

Organizational measurements must also be aligned to create the optimal environment for fighting fraud. Effective anti-fraud training is critical for the success of any fraud detection program, and much depends on the adjuster workflow for detecting and referring suspicious claims. The appropriate integration of process and technology is a must.

The third component of effective fraud-fighting is the development and use of solutions that can identify suspicious claims systematically. In the remainder of this article, let's focus on current and emerging analytical innovations that allow us to build more effective tools for systematic and semi-automated detection of suspicious claims.

Systematic Detection

Building effective fraud-detection solutions requires a strong partnership between business experts, such as investigators and adjusters, as well as analytical staff, including predictive modelers and data analysts. Given the ever-changing nature of fraud, this process is necessarily a virtuous cycle of continual development and improvement.

Until recently, a simple special investigation unit (SIU) scorecard has been the tool for systematically detecting suspicious claims. Typically composed of expert-determined red flags such as indication that the accident was a set up, the claim was reported more than 20 days after loss, or there were unrelated claimants with same doctor, the scorecard has points, or weights, associated with each red flag. The adjusters determine which ones apply to the claim, total the associated points, and refer the claim to the SIU if the total exceeds a preset threshold.

While such scorecards serve a valuable business purpose, they can be further augmented with data-driven predictive modeling techniques. When presented with a historical body of claims, supervised techniques, such as regression models, decision trees, Naïve Bayes, and neural networks, and in some instances unsupervised techniques, such as clustering and outlier detection mechanisms, can produce effective models to determine suspicious claims. The features, or predictors, used in such models are often the expert-determined red flags, with their coefficients or weights being determined by the historical data and its correlation to the SIU action.

Beyond Structured Data

While conventional predictive modeling techniques can yield effective fraud-detection models, they only work with structured data such as code and value fields found in claims and policy databases. Date and time of loss, policy limits, insured driver age, claimant vehicle details, nature of injuries, and treatment duration are all examples of structured data. Often, valuable information that can identify suspicious claims may be buried in other data sources such as the adjuster's notes and claim networks.

Text mining is a recent application in extracting insights from claim notes. While claim notes pose special challenges such as typos, nonstandard abbreviations, grammatical issues, sentence fragments, and changing concepts over time, even simple and crude concept extraction can yield actionable insights. Text profiles can be created to look for adjuster descriptions of interesting concepts such as a low-impact accident¸ an accident near a highway exit, and when the claimant waives EMR/ambulance. An adjuster can use a structured field to denote the presence or absence of this concept in the notes section of each claim. Such structured fields can then be fruitfully leveraged by predictive modeling techniques.

Future Capabilities

Networks that relate claimants, attorneys, doctors, service providers, and other entities on a claim contain valuable information not easily accessible in a structured form. Given a particular person of interest, one can identify all claims associated with that person as well as all other entities associated with those claims. This querying can be applied recursively to multiple degrees by then identifying claims and other entities associated with the entities discovered in the previous query. SIU personnel currently use interactive and deepening queries of this kind as a key capability in the identification and investigation of fraud rings.

However, new capabilities are emerging to exploit claim networks in innovative ways.

Networks—and the participants therein—can be measured by characteristics such as density (number of entities or transactions to which a particular entity is connected) and betweenness (the extent to which an entity is directly connected only to those entities that are not directly connected otherwise), among others. A high density might be expected of a medical provider but not for a claimant in auto accidents. Calculating norms for these characteristics by type of entity can help flag abnormally deviant—and therefore interesting or suspicious—entities and behaviors. Tagging claims and entities with their network characteristics results in additional structured data that can be leveraged by the predictive modeling techniques mentioned earlier.

Emerging technical capabilities, including speech recognition, image and video analysis, telematics, and novel data sources, such as license plate readers, Facebook, and Twitter, will play a key role in creating structured data attributes that can significantly enhance fraud solutions in the future.

Facing the Challenge

Claim fraud continues to be a big business—and a big challenge. In addition to corporate culture, training, and process improvements, fighting fraud requires creating systematic and semi-automated tools to identify suspicious claims. Inventive applications of predictive modeling techniques in fraud detection have been growing with success. Text mining capabilities are expanding the scope of these solutions by discovering and integrating powerful insights from claim text. Growing advances in social network mining are unearthing and characterizing claim networks and participants. Because of the ever-changing nature of fraud, new technologies, capabilities, and data will continually need to be harvested, and innovations pursued, to produce better ammunition in the fight against fraud.

Karthik Balakrishnan, Ph.D., is vice president of analytics at ISO Innovative Analytics (IIA). IIA delivers advanced predictive analytics tools to the property/casualty insurance industry.

Thomas Mulvey is national director of claims and SIU solutions at ISO, an operating unit of Verisk Analytics.