A Classification Model of Uninsured At-Fault Parties to Assess the Risks of Their Claim Repayment Behaviors

One problem in running an insurance business is to request claim repayments from
uninsured at-fault parties. For most cases, it seems very difficult to obtain the full amount of
claims repayment as they avoid or reject to do so. Also, making prosecutions in court can
create additional expenses and may not be worthwhile. This leads to large amounts of losses
for the company and these losses tend to increase yearly. Therefore, the objective of this
project is to find an appropriate classification model of these at -fault parties to predict their
claim repayment behaviors. The data used are the claim repayment follow-up results from
these parties of an insurance company. This consists of a total of all 11,493 cases collected
from the year 2017 to 2022. The methodology and its result are as follows: 1) Find factors
affecting the repayment behaviors (pay/not pay) of these parties by comparing several models.
It is found that the logistic regression model provides the highest efficiency (AUC of ROC
curve = 0.6605) and its 8 factors are Region, At-fault party’s car type, At-fault party car brand,
At-fault party age, Service Type, Claim amount, Covid-19 situation and the CPI, 2) Clustering
all at-fault parties by using the “Proportion of Claim Repayment” and the “Step of
Repayment” as clustering variables, it is shown that these parties can be clustered into 3
groups which are G1: Easy Easy, G2: Pay Only Some Proportion, But Fast, and G3: Never
Pay, and 3) Using all eight factors from step 1) as classifying variables, the confusion matrices
of 5 classification models are compared. It is obtained that the decision tree model with CART
algorithm provides an accuracy of 50.55% and the sensitivities for G2 and G3 are 27.05% and
38.45%, respectively.