A Framework to Forecast Insurance Claims

For an insurance company, the forecasting of claims is central to a successful operation. This process can be divided into multiple subtasks. Data preparation, dimensionality reduction, classification, forecasting and evaluation. This research applies three dimensionality reduction techniques: variable elimination, reduction through a decision forest and multiple correspondence analysis. After dimensionality reduction, classification is used to determine the probability of issuing a claim for an observation to be predicted. Four classification techniques are used: a decision tree, a random forest, a binary logistic regression and a support vector machine. Once the probability of issuing a claim is estimated, it needs to be transformed into a predicted claim amount. As a benchmark, a naive model, called the ratio model, is used. This model uses ratios of risk groups with respect to the base premium to determine the final premium. For the evaluation of the models, classification measures, error measures and the normalized Gini coefficient are used. The results show that dimensionality reduction is not necessarily needed for this problem and that simple techniques, such as a decision tree or random forest, outperform the more statistically advanced techniques, such as a support vector machine, on out-of-sample results.