Neural network models in predicting insurance insolvency and detecting insurance claim fraud.

摘要

Operations research is a quantitative method constructed for solving practical problems in a variety of areas such as logistics and transportation, job scheduling, and strategic management. Operations Research has also enjoyed numerous successful applications in finance. Risk management and insurance is one of the major fields in finance which presents many interesting and challenging issues requiring quantitative thinking and computational solving. Theoretical studies and practical solutions in applying operations research methods to problems in risk management and insurance have been successfully delivered in the past forty years. In this thesis, we provide our overview of many of those studies and applications in effort to shed some light on the future success in the field as well as provide an educational tool.; Neural networks is a branch of artificial intelligence studies. Kohonen's Self-Organizing Feature Maps is one of the major neural network models which have found successful applications, while feed-forward and back-propagation neural networks are apparently the most commonly used neural network model. Both types of neural network models are found to be useful tools in attacking individual problems arising from risk management and insurance. Specifically, one feed-forward neural network is used to predict the insolvency of Texas Property and Casualty insurance companies. This methodology is found to outperform discriminant analysis, logistic analysis and some other rating methods in their prediction accuracy.; Bodily injury (BI) and personal injury protection (PIP) are two major automobile insurance business coverages which suffer serious problems of claim buildup and fraud. We develop a methodology which is a combination of modified Kohonen's Feature Maps, feature map partitioning, and utilization of partially available priori information for the purpose of tackling claim fraud problems in the above mentioned automobile insurance coverage areas. The validation by feed-forward neural networks, used as an approximation, shows that our methodology outperforms the assessments by claim professionals in the consistency of evaluating the suspicion level of insurance claims, and in the quality of result presentation and interpretation. The data sets in our study were provided by the Texas Department of Insurance and the Automobile Insurers Bureau of Massachusetts.