Improving Stability of Tree-Based Models

Abstract

Motivated by the instability of tree-based methods experienced by the insurance industry, this thesis provides innovative and novel methods for updating regression trees in a stable manner. All methods are shown to increase stability, with some methods increasing both stability and performance compared to the baseline. These methods can be extended to random forests and GTB and show similar results as for regression trees. Furthermore, this thesis demonstrates the potential of the stable update methods in improving claims frequency estimation in the insurance industry, but further analysis is required to determine their impact on the premium portfolio.