Tree-based Machine Learning Models with Applications in Insurance Frequency Modelling

Abstract: As the insurance industry is highly data driven it is no surprise that machine learning (ML) has made its way into the industry. While GLMs are still the comfort zone of most actuaries, we have in recent years seen a surge in machine learning algorithms. This study puts focus on developing and evaluating three tree-based machine learning models, starting from simple decision trees and working up to the more advanced ensemble methods random forests and gradient boosting machines. We predict the claims frequency for an all-risk insurance tariff through a case study based on a data set provided by a Swedish insurance company. The gradient boosting machines and random forests are found to outperform the single decision trees, and moreover, we use visualisation tools to uncover and gain insights from the models.