The benefits of machine learning (ML) are widespread. It can help organizations get the most out of raw data, and its applications hold the promise of improving lives. The actuarial profession shares in the potential power of ML, but new technologies can be disruptive. ML actuarial models, for example, can be so complex they lack transparency, threatening to weaken consumers’ trust and raise regulators’ suspicions.
The timely topic of actuarial ML was the focus of a North American Actuarial Journal (NAAJ) article that was chosen as the best in 2021. This prize-winning paper looks inside ML pricing models to understand how they work and explore real-world applications.
The NAAJ’s Annual Prize
Each year since 1983, the NAAJ editorial board selects an article that stands out from the dozens published. The 2021 prize-winning article, “Boosting Insights in Insurance Tariff Plans with Tree-Based Machine Learning Methods,” authored by Roel Henckaerts, Marie-Pier Côté, Katrien Antonio and Roel Verbelen, looks at the inner workings of ML pricing models and develops use cases for pricing actuaries.
Disassembling the Black Box
Are there ways to use ML modeling to improve generalized linear models (GLMs)? Can methods be developed that make ML pricing algorithms understandable to stakeholders? These are a couple of the questions the 2021 NAAJ article of the year investigates.
The authors began collaborating when Marie-Pier Côté, assistant professor and educational leadership chair for big data analysis in actuarial science at Laval University, Quebec City, visited KU Leuven in Belgium for a research stay with Katrien Antonio, a professor in actuarial science and insurance analytics. Both Roel Verbelen and Roel Henckaerts were Antonio’s graduate students at the time. The four of them shared an interest in ML applications, and this brought them naturally to their paper’s topic.
“My team members and I wanted to write a paper that introduced ML concepts in combination with practical use-cases,” stated Henckaerts, now a senior data scientist at Prophecy Labs in Brussels.
The paper provides an overview of tree-based ML methods and provides a case study to show how to adopt ML for actuarial modeling. It follows up with an in-depth comparative study. The authors hope the paper helps actuaries understand how to extract insights from models and assess economic value when comparing it with their existing premium structure.
“The paper brings tools and technicalities together,” said Antonio. “It brings the full story: from the technical details relevant to tune the tree-based machine learning models, to the creation of a technical pricing model, the extraction of essential insights and a thorough comparison with the price list obtained with predictive models (like GLMs) that are well known to actuaries.”
The article’s authors also devised a way to use well-known techniques to explain a model’s logic. This is an important step in creating more transparency. When a model is transparent, consumers understand their risk assessment and resulting pricing structure, regulators can assess the fairness and affordability of premiums, and insurers can assess impact and employ better governance of models.
“Models carry a societal responsibility in terms of fairness and a right to explanation,” explained Verbelen, a senior consultant at Finity in Sydney.
The team has discovered that the techniques they demonstrate in their paper are applicable across insurance products. “The particular field matters less than the specific use case and the data available for it,” stated Verbelen.
In his job at a large consultancy, Verbelen regularly employs tree-based ML methods for general insurance applications, and he uses modeling interpretability tools to understand the models’ reasoning.
Additionally, Côté has seen many insurance companies in Canada use the gradient boosting techniques discussed in the paper as their preferred ML method for claim prediction. And Henckaerts, who has worked in various industries as a data consultant, applies the paper’s techniques in a range of uses, such as determining medical diagnoses, stock market returns, real-estate valuations, fraud detection and churn analysis, to name a few.
Explore Tree-Based ML Modeling
ML actuarial applications continue to grow and affect the profession, and it’s crucial to have techniques and tools that ensure models are working appropriately and transparently. “The most interesting part of being an actuary is thinking through whether models make sense and extrapolate well,” said Verbelen.
Learn more about actuarial tree-based ML models in the prize-winning NAAJ article: “Boosting Insights in Insurance Tariff Plans with Tree-Based Machine Learning Methods.”
Statements of fact and opinions expressed herein are those of the individual authors and are not necessarily those of the Society of Actuaries or the respective authors’ employers.
Copyright © 2023 by the Society of Actuaries, Schaumburg, Illinois.