Predictive Modeling in Automobile Insurance
Award year: 2004-2005
This project will apply the Data to Knowledge (D2K) systems developed by NCSA to the Detail Claim Database (DCD) of the Automobile Insurers Bureau of Massachusetts to generate predictive models to enhance insurance claim investigation practices. Data mining is a relatively new tool for insurance companies. However, advances in applications of data mining have been hindered by the lack of research that can be shared within the industry. Although insurers have conducted many data mining projects for a variety of applications, including underwriting, rating and claim investigation, they have generally resisted disclosing the details of this research in an attempt to maintain a competitive advantage. In many cases, when insurers have tried to utilize the results of these studies in their operations, they have encountered regulatory resistance due to a lack of full disclosure of the supporting documentation. This project will seek to redress some of the problems limiting the advance of data mining and predictive modeling in the insurance industry by conducting a study that can be shared within the industry and all the details can be published. This study will provide a significant advance over the few prior studies that used this data set by utilizing the state-of-the art data mining tools developed at NCSA.
This project will examine the data set for patterns of claim behavior based on the records of the over 1 million automobile bodily injury claims included in the DCD. Both the original D2K system and the more user-friendly D2K Light system will be applied to compare the results and the usefulness of the systems for insurance practitioners. After an analysis of the patterns determined in the data to establish which factors can be effectively utilized in the claims process, a predictive model based on these factors will be developed to help insurers identify which claims are most likely to generate cost savings by investigating the claim more extensively in an attempt to deter fraudulent claiming behavior. The predictive model will be shared with leading insurance companies and consulting firms to gather their input and to provide them with the opportunity to test this model in their operations. Any results obtained from these tests will be incorporated into the publications of this research.