AGGREGATE LOSS PREDICTION USING MULTIPLE-CLASS CLASSIFICATION TECHNIQUES

Abstract

In this thesis, we consider a model for predicting future losses in car insurance applications by using multiple-class classi cation algorithms. In the car insurance payment data, the records can be divided into four di erent groups based on the di erent types of payments, such as labor-dominant (LD), parts- dominant (PD), other-dominant (OD), and none-dominant (ND) payments. We rst apply the multi-class classi cation algorithm to predict the probabilities that a ran- domly selected subject belongs to the corresponding group, then, for simplicity, the predicted payment amount can be calculated by using the total expectation formula after determining the conditional expected payment amount using training data. This method is easy to implement and yield satisfactory prediction accuracy. We compared proposed model accuracy against general linear regression and clas- sical individual aggregate loss models accuracy for the test dataset. The comparison results show that the proposed model outperforms other models in accuracy for given test datasets. The matrix used to describe the distribution of the true groups and the distribution of the predicted groups indicates there exists a relationship between the payment amount and the type of group it belongs to.