EAA Web Session: "Building an Interpretable Machine Learning Model"

Einleitung/Dauer

In the face of the ever-increasing competition and mounting regulatory pressure, actuarial precision and accuracy shape the art of setting the price in the non-life insurance sector. Generalized Linear Models (GLM) are the standard pricing method of non-life insurance products, leading to a multiplicative tariff that is immediately interpretable and operationally efficient. In recent years, the advent of Machine Learning has been termed the next frontier of innovation and productivity, focusing on prediction performance and capturing the inherent non-linearity of the data. However, there is a need to associate these complex models with interpretability techniques.

We introduce an Explainable Boosting Machine (EBM) model that combines both intrinsically interpretable characteristics and high prediction performance. This approach is described as a glass-box model and relies on the use of a Generalized Additive Model (GAM). In this web session, rather than explaining Machine Learning models, we aim to build models that are intrinsically interpretable.

Firstly, we recall the parametric structure of the GLM model as well as the non-parametric structure of Machine Learning models. Some general principles of Machine Learning methods are also presented such as model aggregation.

Secondly, we focus on the GAM model and its declinations. We present its semi-parametric structure and the smoothing-learning paradigm within the shape functions. The EBM model is then introduced as an example of GAM combining Machine Learning functions.

The third part gives a general overview of interpretability techniques and aims to position the EBM interpretability among them.

Lastly, we provide an application of EBM using a Jupyter notebook designed around a P&C actuarial use case.

Vorgehensweise und Ziele

The purpose of this web session is to propose a model at the frontier of GLM and Machine Learning methods. By the end of the web session, participants will understand mathematical principles behind the EBM model, assess its interpretability and apply it in an actuarial framework.

Teilnehmer

This web session is intended for P&C actuaries, statisticians and data scientists who are interested in Machine Learning techniques and their applicability to the insurance sector. A basic knowledge of Machine Learning concepts and some programming skills (e.g. Python or R) are a plus.

Technical Requirements
Please check with your IT department if your firewall and computer settings support web session participation (the programme Zoom will be used for this online training). Please also make sure to join the web session with a stable internet connection.

Dozierende

Marketa Krupova
Marketa Krupova has an engineering degree in applied mathematics from the National Institute of Applied Sciences and a master’s degree in actuarial sciences from the Paris Dauphine-PSL University. She is an actuary fellow of the French Institute of Actuaries and currently works at the consulting firm Addactis with a focus on pricing and data science in non-life insurance.

Sprache/Kurztitel

The language of the web session will be English.