Operationalizing a Fraud Detection Solution: Buy or Build?
James Ruotolo
With challenging economic times driving the need for better underwriting discipline and expense control, predictive analytics can provide an advantage. Fraud detection is a common use-case for early adoption of analytics within an insurance organization because the return is high, the risk is relatively low, and claims data tends to be robust enough to build effective models. In fact, this phenomenon is starting to change the face of Special Investigation Units (SIUs) in the industry. "We used to hire ex-military and police, but now we're hiring modelers [in our SIU]," commented Frank Llende at Allstate during a recent panel discussion.
[For more on James Ruotolo's insights on the application of analytics to the fraud challenge, see Direct to Fraud: Applying Analytics to the Fraud Challenges of the Direct Insurance Channel.]
With many insurers pursuing an analytical fraud detection solution, a common question arises: whether to buy a solution or build one internally? Larger insurers typically have top analytical talent within their organizations, with small armies of actuaries, statisticians and predictive modelers. Undoubtedly, major insurance companies have the capability to build effective predictive models. Homegrown solutions allow total control over development. The temptation is great to build a fraud scoring model. However, more and more companies are rethinking the internal approach and evaluating external solutions for several key reasons.
Fraud detection is more complicated than it seems
The best fraud detection models incorporate multiple fraud detection techniques – not just business rules or supervised predictive modeling. This requires expertise in areas like text mining and social network analytics that may be unfamiliar to traditional insurance modelers. Also, standard data quality approaches often used by insurers need to be modified for fraud detection purposes or else it is possible to wipe out the anomalies that are indicative of fraud. In addition, fraud detection models must be tuned periodically – usually one or two times per year at minimum – to maintain peak performance as fraud schemes constantly evolve.
Building models is only half the battle
Once the models are built, they need to be put in production. Organizations quickly discover that end users – usually analysts or investigators in the SIU – need a way to review the suspicious claims, route alerts to key resources, and capture user feedback as the alerts are disposed. Emailing spreadsheets across the organization becomes unmanageable very quickly. A proper implementation requires the creation of user interfaces, new databases and metadata to capture feedback, regular reports on model performance, and administrative tools to control everything. If an insurer builds these capabilities, the efforts to operationalize a model can add 12-18 months onto a project.
Total cost of ownership
It is easy to underestimate the infrastructure and effort required to deploy and maintain a robust solution. Once all the costs and effort for building a production environment are evaluated, the business case for a purchased solution becomes much stronger. While insurers may have the staff to build their own predictive models, they often discover that the total cost of ownership for a fraud detection solution is lower with a vended solution.
Internally-built solutions and vendor-provided solutions each have many benefits. Insurers should carefully consider both options when looking to deploy anti-fraud technology. Regardless of whichever route your organization chooses, analytical fraud detection systems will help insurers detect suspicious activity faster and with greater accuracy than ever before.
About the Author: James Ruotolo is an insurance fraud technologist, thought leader and the principal for insurance fraud solutions at SAS. Connect with him on Twitter or LinkedIn.
Tweet This