Imbalanced learning using actuarial modified loss function in tree-based models

The point mass at zero and the heavy tail of insurance loss distribution poses the challenge to apply traditional methods directly to claim loss modeling. Via an illustrative simple dataset, this thesis first pinpoints the pitfall in the traditional tree-based algorithm’s splitting function. This thesis then modifies the function to remedy the imbalance issue presented in the insurance loss modeling. We propose two novel actuarial modified loss functions, namely, weighted sum of squared error and Canberra loss functions. This modification imposes a significant penalty on grouping nonzero observations with zero ones at the splitting procedure. We examine and compare the predictive performance of such actuarial modified tree-based models in relation to the traditional models in a synthetic dataset. Our studies show that, such modification results in improved prediction and completely different tree structures.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Actuarial Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2021

Author dc:creator

• Hu, Changyue

Contributors dc:contributor

• Quan, Zhiyu
• Chong, Alfred

• tree models, imbalanced learning, predictive modeling

Statement dc:rights

• Copyright 2021 Changyue Hu

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/110877

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/110877