Applying Machine Learning to Detect Automobile Insurance Claims without Surveys

Abstract

Automobile insurance was the most fraudulent because claims without surveys, also known as “dry claims,” led to fraud and arose from the insured failing to provide details of the incident match with reality. Traditional fraud detection methods were difficult to administer. This research utilized machine learning to solve the above problems. The dependent variable used in the study was compensation claims (dry or fresh claims). The research objectives consisted of 1) searching for factors influencing compensation claims; 2) comparing the predictive performance of the three variable selection methods, no selection of independent variables, Stepwise logistic regression, and decision tree; 3) comparing the predictive performance of the three algorithms, namely Naïve Bayes, random forest, and adaptive boosting; and 4) comparing the predictive performance of the algorithms together with variable selection methods. The claim report of the Type I automobile insurance from an anonymous insurance company in Thailand in 2021 was employed in this study. The study found that the cause of the incident, scene region, insured age, type of repair, sum insured, and amount paid influenced the claims. The most effective variable selection method was the decision tree. Random forest, adaptive boosting, and Naïve Bayes was the most efficient algorithm, respectively. The random forest and the independent variable selection method with the decision tree method were the most efficient.