A multi-task network approach for calculating discrimination-free insurance prices (Q6593136)

The paper focuses on the intriguing question of avoiding discrimination in insurance pricing. A multi-task neural network architecture for claim prediction is used to avoid indirect discrimination in insurance price calculation. It is possible to train the system with only partial information about protected characteristics and produce prices without any form of proxy discrimination. After presenting the key concepts of the discrimination-free insurance pricing, a solution to the problem of incomplete protected information is provided, gradually moving towards the multi-task network architecture. The methodology's performance is demonstrated by a synthetic health insurance example and the suggested approach is applied to a real-world non-life motor insurance dataset.