What do a Harley, a Ferrari, a fire, and a fake accident have in common? A policyholder—we’ll call him Frank—submitted a claim for the theft of his Harley Davidson Street Glide to his current insurance carrier. This isn’t unusual; he claimed thieves used a trailer to steal the motorcycle while parked in front of a friend’s house. Unfortunately for Frank, his insurer was using advanced technologies powered by industry data to help determine which claims to pay quickly and which to investigate.
Fraud is constantly evolving—as are the means of fighting it. To better understand the future of fraud, it’s important to understand rule-based scoring, easy customization, and AI-driven predictive models.
In Frank’s case, his insurer’s predictive model assigned a high fraud propensity score to his claim and provided information to investigators as to why.
Evidently, Frank had a significant history of submitting diverse types of insurance claims to multiple insurance carriers over many years. He’d also previously been charged with two counts of theft. Additionally, Frank carried a considerable amount of bad debt on various credit cards turned over to collections, and his home lender had filed a foreclosure notice for delinquent payments.
The insurer’s investigation also uncovered a prior claim Frank had made for a “fix and flip” house a few months after he purchased and insured the property. During the remodel, a fire broke out—reportedly caused by stained rags and spontaneous combustion in an enclosed area. Unfortunately, all the remodel work burned, and the property damage amounted to over $100,000.
Many years before this, Frank had owned a red Ferrari Testarossa that had been previously wrecked and had incurred extensive damage. So, Frank purchased it for a “good deal” and made all the necessary repairs to restore the car. However, it was later reported stolen while Frank was dining at a high-end steakhouse after a valet parked the car. Interestingly, the video camera in the valet lot was not functional at the time of the theft.
Frank was also previously involved in a major T-Bone accident at a rural intersection when a stranger, Mr. Miller, ran a stop sign. Frank was in a truck and claimed soft tissue, TMJ, and head injuries when his truck was totaled. The alleged ‘stranger’ was very apologetic in his statement to the police and claimed to be driving back roads on his way to visit his friend’s farm close by in the neighboring state.
The predictive models employed by Frank’s insurer scored his claim based on this diverse history of suspicious claims and financial motive information available, prompting an investigation that could save the insurer a substantial amount of money and bring a fraudster to justice.
Successfully fighting fraud, as the methods of committing it evolve, means leveraging artificial intelligence (AI)—utilizing predictive modeling specific to a line of business and trained on broad industry data—in parallel with easily customized, carrier-built rules. Integrating these tools into a hybrid approach is the strongest perimeter defense against fraud.
Other capabilities are emerging in the industry as well. Advances in digital media forensics allow for the detection of fraudulent images. It’s now possible to recognize in-database duplication, internet duplication, and altered metadata. This field is expanding to include recognition of altered pixels and documents. Additionally, the ability to score medical providers and their treatment continues to advance, using advanced analytics to identify anomalies and outlier trends—resulting in a significant reduction in manual lift. As fraudsters advance, so do the available nets of detection.
AI Predictive and Machine Learning Models
In the quest for slowing down questionable claims for investigation and allowing fast pay for others, predictive models with machine learning capabilities are trained on data to identify which claims have the potential for fraud. With experience, they can be retrained from claim triage and disposition results to refine and expand fraud recognition and surface questionable claims.
AI-driven models score claims and consider each variable and their interactions. Although business rules are important, they are both rigid and static, requiring human configuration. Conversely, machine learning systems learn to recognize complex patterns from the data feeding the algorithm, making them helpful to optimally prioritize fraud risks for review. Using a hybrid approach proves helpful.
The Hybrid Approach
While predictive models are powerful tools for deducing complex patterns in data and predicting fraud, it is also important to recognize one of the primary limitations of a predictive model in the context of fraud. Fraud is not only relatively rare overall, but fraud patterns themselves adapt and evolve over time.
Predictive models require targets to learn new patterns, and the necessary accumulation of evolving target data does not happen overnight. At best, this creates a lag in when the model effectively starts identifying an emerging fraud scheme. At worst, the pattern doesn’t get detected, given that machine learning systems depend entirely on the data fed into them.
A strong perimeter defense requires diversifying analytical techniques and tools to identify fraud. A flexible and efficient rules engine and architect solution can help fill the algorithm learning gap by supporting the implementation of custom fraud scenarios on the fly—to quickly respond to newly discovered emerging threats or otherwise experiment with new means of generating fraud leads. Over time, this provides your AI-driven model with more inputs and targets to improve.
Let’s talk about older and new rule systems—those which might be considered dated and the changes that are important to understand.
You will find rules in every vendor fraud technology solution on the market today. A rule-based system works with “if-then” algorithms established to determine simply whether a particular situation exists: a “yes,” or “no.” Rule-based systems are binary. The situation the rule describes is either present or it isn’t. In contrast to trainable predictive models, a rule does not learn. And, by design, it doesn’t need to. The value of a rule can be understood through a few examples:
- A rule might identify if an involved party had a prior SIU claim that either wasn’t paid in part or was denied altogether. Knowing this information can be essential. That said, the answer is binary. It is either yes or no.
- A rule might recognize if an involved party has had three or more prior water claims—certainly a suspicious anomaly. The answer is binary: either yes or no.
- A rule might recognize an involved party with two prior hit and run accidents, a parking lot accident, and two prior thefts. Binary results by nature? Yes. Of value? Of course.
These examples demonstrate key decisioning information. Knowing the answers to these rule-based questions will be as important five years from now as they were five years ago.
Beyond rules, predictive analytics generate significant opportunities in parallel to expand detection and recognize the complexity and diversity of identifying fraud. Therefore, a hybrid approach that includes rule-based systems and predictive modeling and scoring increases the likelihood of surfacing questionable claims with optimal operationalization.
Data Is Key
The right data is critical to detecting whether these important scenarios are present. The effectiveness of rule-based systems is determined by the data (or lack thereof) needed to distinguish whether something significant is or is not present—whether it did or did not occur.
For example, if a carrier’s optics do not include a complete claim history generated by data aggregated at the industry level, then the results will either be limited or missing.
When it comes to data, the most comprehensive source in the industry is Verisk’s ClaimSearch database, which now holds almost 1.7 billion claims and grows daily. The power of historical context, which also applies to fraud, dictates that the past informs the future. Therefore, the ability to identify years of claim history for involved parties, vehicles, locations, and more is paramount to connecting relevant and current associated claim details. And the recognizing histories across insurance carriers—regardless of the policyholder at the time—is especially important.
The ability to easily and quickly self-administer business rule configurations is an industry-leading practice. Fraudsters change with time. Their schemes evolve, which necessitates engagement with predictive analytics and machine learning and leveraging the newest business-rule-based solutions to function as an architect—designing rules and variables within a self-administered system.
Without easy self-administered customization, the alternative is finding the budget for expensive and hard-to-schedule IT resources to configure rules further or dealing with a lack of flexibility. Both alternatives lead to diminished results, timeliness, effectiveness, and inventory quality—all while manipulative fraudsters continue to develop their craft.
Easy business rule customization with the ability to create your own rules in a sandbox environment and roll to production when and where you see fit is critical to fostering the intelligence and creativity of your fraud experts. But the key is ensuring this end-to-end rule customization and architect capabilities can be performed with minimal IT lift to avoid massive delays and hidden costs.
A Full Defense
Verisk’s Anti-Fraud One solution is a full-service, hybrid approach that provides carriers with:
- AI-driven, predictive modeling with machine learning
- Full-service business rules with easy and inexpensive self-administration
- The ability to architect and self-design needed rules
- A full suite of Image Forensics
- Options for medical provider scoring and rationale
- Powerful aggregated industry data for claims and medical providers
- Third-party data sets, including civil, criminal, and financial motive data
- Consultation from subject matter experts who understand fraud investigation, leadership challenges, and anti-fraud tech solutions
Claims might look legitimate. A fraud-committing claimant – like Frank in our earlier example—might even be friendly, intelligent, and seem helpful. But appearances can be deceiving. Predictive modeling can automatically and immediately surface alerts to trigger an investigation to help prevent payment for questionable claims to questionable people.
You do not have to lose months and incur exorbitant costs extracting historical data sets, scheduling IT support, and mapping all your data to a vendor to test and evaluate a predictive model solution. This massive undertaking inevitably creates additional ripples across departments and generates unknown costs.
Verisk offers a unique solution that will enable you to onramp a pilot in about one week, partner with a fraud expert, and test our solutions. Quickly evaluate the impact of advanced analytics in your operation and determine a solution’s value and return on investment.
What happened to Frank? In the end, Frank’s claim history combined with his current financial stressors/motive was too much to overcome. Frank’s claim history with multiple insurance carriers was thoroughly investigated as a part of his ongoing claim. While inspectors initially cleared the fire damage to the remodeled house, a statement was obtained from the friend helping Frank remodel his house—a friend with the last name Miller. Yes, the same “stranger” involved in Frank’s T-Bone accident. The carrier’s Exam Under Oath attorney indicated arson was likely. A connection to the stolen Ferrari parts was also discovered. Law enforcement located the stolen Ferrari doors in another state and tracked them back to an out-of-state salvage company, coincidentally owned by the uncle of Mr. Miller.
It was ultimately determined the Harley was an owner give up by Frank for insurance money. His claim was denied and reported to authorities. Frank was convicted of insurance fraud and sentenced to jail. AI-driven predictive analytics—like those that power Verisk’s anti-fraud solutions—help insurance carriers prioritize investigations, unravel a web of hidden schemes like Frank’s, and avoid claims without merit.