Design and Development of Fraud Detecting Tools and Scenarios for the Insurance Market

The discourse with the insurance anti-fraud units and the continuous attention to the needs of the client have been placed as central in the development model of Shift, in order to analyze and improve the results produced by company's fraud detection system. To best implement this business strategy and respond naturally to client needs, multiple teams of data scientists are part of the Shift organization and interact in a native way with the insurance company's anti-fraud team to understand its needs and guide tool development. In this perspective, the results of the consulting activity and the improvement on the anti-fraud tool will be described, based on discussions of the operational needs and a bi-weekly analysis of the performances. For the implementation of the planned actions, it will be essential to understand the structure of the project and the different modules that compose it; as well as the understanding of the complex relational data schema of the company, core of the Shift’s second-generation detection system, whose details will be described in a special section of this work. The still ongoing pandemic has altered our lifestyles and so has the typology of frauds that are concocted. The result of the activity of consultancy, of the duration of approximately five months, will consider the introduction of a series of changes that will require to design, build and develop new detection features and then integrate them in the Fraud Detection tool. The most important part of the work will focus on identifying new fraud contexts, born during the pandemic, through the design and development of new fraud scenarios or the restructuring and calibration of existing ones. The changes made will then be measured through performance indicators, analyzing how they affect the detection of possibly fraudulent claims. Another effort will concern the complete automation of the tool and the statistics it produces, through the creation of special jobs that will reduce the human tasks required to a minimum, granting a more regular and responsive communication of key values to the client. In order to fully exploit data sent daily by insurers, we will then design a solution for the interpretation of the free text fields entered by the user in the description of online claims. A technique based on natural language processing will finally be proposed; potentially adaptable to any context due to the high modularity and flexibility of the solution, it will be applied to poor, highly-variant and unbalanced datasets. This new model will be able to improve the the recognition of potentially simulated diagnoses and deny payments for staged trip cancellations, helping the client to save large refunds. The result of these actions, in order to improve the performance of the tool for the client, will contribute to the saving of a significant amount of money due to fraudulent claims and increased number of additional suspicious cases under investigation, as well as the consolidation of business relations between the two companies.