Variable Selection with Machine Learning in the Legalization Process for Traffic Insurance

Abstract

In the insurance sector, the insured notifies the insurance company of which he is the customer as soon as the damage occurs. Upon this notice, a claim file is opened to the insured, and the damage file number is assigned. The claim file contains information about the product insured by the insured and the damage. This information is kept in tables in the databases of Anadolu Insurance. In the event of damage, the insured's claim can be accepted. The entire damage amount can be paid if the damage amount is partially accepted, with the examination to be carried out by the insurance company; if the damage amount is partially accepted, a part of it is paid, or the claim is rejected. The damage amount is not paid at all. When the Insured receives partial payment or the claim file is rejected, they can sue the insurance company to claim the damage amount. The litigation process is a long and bad experience for the insured. For the insurance company, in addition to customer dissatisfaction, it causes extra costs such as court, lawyer, etc. costs. The problem studied in this work is aimed to determine which variables are essential for a possible legalization process in case of partial acceptance or rejection of the claim file by using the variables in the relevant claim file by machine learning and statistical methods. While making this determination, lasso regression, information gain, chi-square test, fisher's score, Recursive Feature Elimination (RFE) with Random Forest Machine Learning algorithm, Univariate Feature Selection with bivariate statistical tests or univariate statistics like chi-square test and feature importance of Random Forest Machine Learning algorithm. Variable selection was made by using correlation coefficient and backward feature elimination methods. Variable p_value was also evaluated.