SHAP values (Shapley additive explanations)

Shapley values come from cooperative game theory: in 1953 Lloyd Shapley proposed a unique way to share a coalition's gain among its members, according to each one's average contribution to all possible coalitions, while respecting fairness properties. The SHAP method, published in 2017, applies this principle to explaining predictive models: for a given prediction, for example an insured's premium or a claim's fraud score, it computes each variable's contribution to the gap between that prediction and the average prediction. The contributions sum exactly to that gap, which makes the explanation consistent. Aggregated over a portfolio, the values show each variable's overall influence and the shape of its effects. The method has become the standard for interpreting ensemble models and neural networks in insurance, notably to meet regulators' explainability requirements and to justify a decision to a policyholder. Its limits lie in computational cost, sensitivity to correlations between variables, and the fact that it explains what the model does, not why the risk is what it is: a discriminatory proxy variable appears in it like a legitimate variable.