Fredrik Käll: Gradient Boosted Trees Applied to Chain-Ladder Reserving
Time: Wed 2022-09-07 13.45 - 14.30
Respondent: Fredrik Käll
Supervisor: Mathias Lindholm
For information on how to enter the department, send an email to firstname.lastname@example.org.
This thesis investigates whether non-life claims reserving can be im- proved by using more information regarding each claim and machine learn- ing techniques. As a foundation, Wühtrich’s article Neural Networks ap- plied to Chain-Ladder reserving has been used, with the modification that Gradient Boosted Trees have been used instead of Neural Network. We begin by obtaining a model by walking through the fitting process. The model is then used to predict the outstanding reserves and compared to Mack’s Chain-Ladder predictions. Further, a comparison of the two mod- els MSEP is made to investigate the variation of the two models. The comparison shows that the Gradient Boosted Trees perform as well as Chain-Ladder for earlier, more developed years. However, in later years, the performance is not as good. We end the thesis with a discussion on why the boosted trees did not perform well and what could be done to improve the predictions.