Group Shapley values, with application to bond recovery rate prediction

Grouped feature attribution is increasingly important in high-dimensional applications, where predictors often belong to economically or scientifically meaningful categories and individual SHAP values can become too granular for reliable interpretation. This paper develops a formal inferential framework for assessing whether group-level SHAP-style attributions are statistically significant and complements it with a tree-based grouped implementation for ensemble models. The proposed testing procedure is based on a three-cumulant chi-square approximation and is designed for challenging settings involving high dimensionality, skewness, sparsity, feature dependence, and limited sample size. Simulation results show that the proposed test maintains accurate empirical size and often achieves higher power than competing alternatives such as the Wald and CQ tests. The framework is applied to bond recovery rate prediction using a global dataset from 1996 to 2023 with 2094 observations and 98 features organized into 16 subgroups and five broader categories: bond characteristics, firm fundamentals, industry-specific factors, market-related variables, and macroeconomic indicators. The empirical results identify market-related variables and bond characteristics as the most influential groups. Overall, the paper contributes a statistically rigorous approach to global inference on grouped SHAP-style importance measures and illustrates its practical value in a financial prediction setting.