Customer profitability forecasting using fair boosting : an application to the insurance industry

Authors: St-Jean, Alex
Advisor: Marchand, Mario
Abstract: Customer profitability forecasting, along with ratemaking, are central pieces in the world of actuarial science. By using historical data and by optimising statistical models, actuaries can predict whether a client with certain liabilities will claim any loss and what amount will be claimed inside a defined policy period. However, these models often use sensitive attributesrelated to the customer that are considered to be crucial risk factors to consider in predicting future losses. This is considered legal in many jurisdictions, as long as their use is backedby actuarial data, as these attributes give a more accurate premium to clients. Nonetheless,as it has been noted in recent machine learning literature, models can hide biases that makethem discriminate against certain groups. In this thesis, we propose a customer profitability forecasting model that uses recent advancements in the domain of machine learning to ensurethat these algorithms do not discriminate disproportionately on a subgroup of any intersectionof protected attributes, such as age, gender, race, religion and marital status. In other words,we fairly predict the theoretical premium of any client by combining state-of-the-art methodsin insurance loss prediction and the application of fairness constraints on regression models. After performing customer profitability estimation on multiple real world datasets, it is shownthat the proposed approach outperforms traditional models usually used for this task, whilealso satisfying fairness constraints. This shows that this method is viable and can be used inreal world scenarios to offer fair and accurate premiums to clients. Additionally, our model andour application of fairness constraints scale easily when using large datasets that contain many subgroups. This can be substantial in the case of satisfying an intersectional fairness criterion.Finally, we highlight the differences between actuarial fairness and fairness definitions in theworld of machine learning, along with its related trade offs.
Document Type: Mémoire de maîtrise
Issue Date: 2021
Open Access Date: 8 February 2021
Grantor: Université Laval
Collection:Thèses et mémoires

Files in this item:
Description SizeFormat 
36951.pdf1.45 MBAdobe PDFThumbnail
All documents in CorpusUL are protected by Copyright Act of Canada.