Restriction lift date: 2030-12-31
Performance and fairness of machine learning modelling for general insurance pricing
Loading...
Files
Date
2024
Authors
Israni, Tarun
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
An increasing number of reports underscore the promising potential of machine learning (ML) methodologies for general insurance pricing compared to the conventional generalised linear model (GLM). This shift towards ML is coupled with a growing emphasis on pricing fairness by national and international regulatory institutions. However, the scarcity of comprehensive studies that assess both pricing accuracy and fairness is a challenge for insurance companies and regulatory bodies alike. We propose a comprehensive study of the GLM - acting as a benchmark model - against mainstream regularised linear models and tree-based ensemble models. We evaluate these models in two common pricing frameworks (Poisson-gamma and Tweedie), considering key criteria such as estimation bias, deviance, risk classification, competitiveness, loss ratios, discrimination, and fairness.
This study includes an evaluation of three definitions of fairness - demographic parity, actuarial group fairness, and calibration - with respect to protected policyholder information at a population level. Pricing performance and fairness are assessed simultaneously on the same samples of premium estimates for each model. The models are assessed on two open-access motor insurance datasets, each with a different type of cover (namely fully comprehensive and third-party liability).
While no single ML model outperformed the others across all metrics, the GLM significantly underperformed for most. The results demonstrate the potential of ML as a realistic and reasonable alternative to current industry practices. The last chapter of this thesis proposes to consider fairness under a different light, focusing on the analysis of ratios of actual to technical premiums. The study incorporates metrics from the Central Bank of Ireland (CBI) differential pricing review, to examine the impact of gender, tenure, and age on pricing. Overall, ML models outperformed GLMs over most metrics for general insurance pricing. However, achieving accuracy and fairness simultaneously is not straightforward, and fairness may be understood in various ways.
Description
Keywords
Machine learning , Actuarial pricing , Pricing bias , Pricing structure , Rate making , Fairness , Protected variables , General insurance , Non-life insurance
Citation
Israni, T. 2024. Performance and fairness of machine learning modelling for general insurance pricing. PhD Thesis, University College Cork.