With newer technologies and information available to the general public, there are substantial automobile insurance frauds occur, causing a huge loss in various insurance companies. To reduce the loss, insurance companies adjust their business strategies and insurance policies, such that increasing premiums for the customers, which alleviate the risk from insurance companies but hurt the benefits of customers. Since customers are unsatisfied with these strategies, the retention rate has been decreased, forcing insurance companies to find a better idea to reduce the fraud cost. To solve this problem, there is no better way than detecting this group of customers who potentially more likely make fraudulent claim and rejecting them at the first place.

The purpose of this research is to develop a predictive model to decide whether an insurance claim is a fraud or not, in order to reduce the future unnecessary cost and potentially protect innocent insurance customers. To achieve the goal of this research, both parametric and nonparametric methods were applied. For reducing uncertainty and increasing the chances of detecting the appropriate claims, we considered statistical learning methods by emphasizing variable selection. The proposed models measured the importance of individual predictors and classified the response as fraud or non-fraud, based on the probability. For parametric methods, we applied logistic regression and regression with LASSO regularization. For nonparametric methods, we applied random forest and SVM models. To learn deeper, we applied the Neural Network algorithm to investigate a deep-learning network. We investigated the performance including the balance between sensitivity and specificity to classify fraudulent claims via the test set.