Individual Model Explanations Using Avatars

Explaining individual insurance premiums is becoming increasingly important in insurance data analytics, as it enables both policyholders and regulators to understand how premiums are determined and ensures model transparency and regulatory compliance. Finite-change Shapley values are a local sensitivity method that quantifies the contribution of individual risk factors to the premium difference between a policyholder and a reference individual. However, estimating these indices from real-world data requires evaluating the pricing model at synthetic policyholders whose feature combinations may not be compatible with those observed in the data, potentially leading to extrapolation errors and misleading conclusions. In this work, we introduce the concept of “avatars”. Avatars are synthetic policyholders for which the feature combinations are compatible with those observed in real data. The use of avatars mitigates the extrapolation problem and offers additional advantages: Since avatars are not based on real individuals, they can be publicly disclosed, allowing policyholders and regulators to assess the transparency of the pricing model without privacy concerns. A Monte Carlo approximation is also introduced to make the computation feasible for large insurance portfolios. Using a medical health insurance dataset, we illustrate how this methodology provides useful insights into the sensitivity of premiums to individual risk factors.