Claim Severity Prediction of Workers' Compensation Using Tree and Neural Networks-Based Models

Along with the increasing trend of work accidents during the 2007-2017 period and the resumption of business activities after the COVID-19 pandemic, workers' compensation insurance has become a potential line of business to be developed in Indonesia. As an essential component in the insurance business model, the claim severity affects the determination of premium rates for the insured. Machine Learning has become a potential alternative method for claim prediction. Because of its popularity, a Transformer-based Neural Networks model is proposed for a tabular data model called TabTransformer. This paper examines the Neural Networks-based models, i.e., Multi-Layer Perceptron (MLP) and TabTransformer, to predict workers' compensation insurance claims and compare their accuracies with some Tree-based models, i.e., Decision Tree, Random Forest, and XGBoost models. Our simulation on three datasets shows that the Tree-based models provide better results than the Neural Networks-based models, and XGBoost is the best one. MLP provides better results than Tree-based methods on one of the three datasets. Moreover, TabTransformer succeeded in improving MLP accuracy only on one dataset.