IA|BE Chair 2022: 20/09/2022 (16h00 - 18h00)
Topic: 'Discrimination-Free Insurance Pricing: From Concepts to Implementation (joint work with Mathias Lindholm, Ronald Richman, Mario Wüthrich)
Speaker: Prof. Andreas Tsanakas, Professor of Risk Management
Andreas Tsanakas joined the Bayes Business School in 2006. Previously he spent six years at Lloyd's.
Andreas' research interests are in quantitative risk management, with particular focus on portfolio risk measurement, sensitivity analysis, capital allocation and model uncertainty.
He is Editor-in-Chief of the Annals of Actuarial Science, Associate Editor of ASTIN Bulletin, co-organizer of the annual Insurance Data Science Conference, co-organizer of the One World Actuarial Research Seminar (OWARS), and co-author of the R package SWIM, used for efficient sensitivity analysis of simulation models.
Given information on individual policyholder characteristics, how can we ensure that insurance prices do not discriminate with respect to protected characteristics, such as gender?
We address the issues of direct and indirect (or proxy) discrimination, the latter resulting from implicit learning of protected characteristics from nonprotected ones.
We provide mathematical definitions for direct and indirect discrimination and introduce a simple formula for discrimination-free pricing, which avoids both direct and indirect discrimination. This formula works for any statistical model and relies on building best-estimates using all policyholder characteristics (including protected ones) and then averaging out the protected characteristics for calculating individual prices.
However, such approaches require full knowledge of policyholders' protected characteristics, which may in itself be problematic. We address this issue by using a multi-task neural network architecture for claim predictions, which can be trained using only partial information on protected characteristics, and produces prices that are free from proxy discrimination.
We demonstrate the use of the proposed model on a synthetic health insurance example and find its predictive accuracy comparable to a conventional feedforward neural network, when full policyholder information is available.
However, this multi-task network has clearly superior performance in the case that information on protected characteristics is partially missing.
We conclude with some comments on the relationship between our proposed approach and considerations of fairness.
This talk is based on the following two papers:
Lindholm, M., Richman, R., Tsanakas, A., & Wüthrich, M. (2022). Discrimination-Free Insurance Pricing. ASTIN Bulletin, 52(1), 55-89.
Lindholm, M., Richman, R., Tsanakas, A., & Wüthrich, M. (2022). A multi-task network approach for calculating discrimination-free insurance prices. Preprint. https://arxiv.org/abs/2207.02799
Some practical information:
One day before the websession, you will receive a reminder with the link needed to participate in the websession. Please log-in via this link 5 minutes before the start of the event.
In case of any question, please contact us.