Abstract:
"In Sri Lanka, one of the main challenges faced by health insurance providers is identifying
future risk of policyholders to come up with competitive and affordable premiums. Risk
is a factor that differ from person to person and careful identification of risk is a crucial
part of underwriting process. With the increase of insurance data and the use of machines
learning algorithms the underwriting process can be further improved by faster data
processing and identification of risks. This research aims at providing a framework
coupled with a machine learning pipeline to predict health insurance risk amount in an
accurate manner. A real-world dataset containing five years of claimed data has been used
to conduct the analysis. The research was carried out by training seven classification
models and four regression models to develop a machine learning pipeline to classify risk
and then to predict risk amount. The experimental results showed that the proposed
machine learning pipeline has obtained acceptable results and ensemble algorithms works
well on health insurance data compared to other machine learning algorithms. In
classification part of the study, Random Forest and XGBoost achieved the best accuracies.
The regression part of the study reveals that XGBoost perform well in predicting amounts
for low-risk policyholders and Random Forest generate better results for normal, high,
and bad risk data after applying data augmentation techniques."
