Higher Order Shapley Effects for Global Sensitivity Analysis of Extremes

Variance-based Shapley effects are attracting increasing attention as a global sensitivity analysis tool for interpreting complex models. The Shapley effects allocate the variance into individual input's contribution and remain well-defined also in the case of dependent inputs. 

In this work, we introduce the Shapley effects for the skweness decomposition. These new indices identify which inputs are more responsible to make the model output attain its extreme values. We provide a given-data computational algorithm and we illustrate our findings using economic and financial datasets.