Addressing unanticipated interactions in risk equalization: A machine learning approach to modeling medical expenditure risk

Adverse selection harms market efficiency and access to essential services, particularly for disadvantaged groups. Risk equalization policies attempt to mitigate this by compensating agents for risk disparities, but often fall short of addressing interactions between risk factors. Using health insurance data from the Netherlands, we present a machine learning approach to capture unanticipated interactions that impact medical expenditure risk. We compare our novel approach to a state-of-the-art statistical model. We find that our approach explains an additional 1.5% of medical expenditure, equivalent to 571 million euros over all individuals in the Dutch market. In particular, this translates into better compensation for low- and high-cost groups that are especially vulnerable to adverse selection. These findings confirm the significance of risk factor interactions in explaining medical expenditure risk, and support the adoption of machine learning alongside statistical models to further mitigate selection incentives in risk equalization policies.