Group Shapley with Robust Significance Testing and Its Application to Bond Recovery Rate Prediction

Group Shapley provides more interpretable AI explanations for structured data: The paper introduces Group Shapley as an extension of Shapley values that evaluates importance at the feature group level rather than individual features. This is particularly relevant for real-world business applications where features naturally cluster into categories (e.g., bond characteristics, firm fundamentals, macroeconomic factors), making AI model decisions easier for stakeholders to understand and act upon.

The method includes statistical rigor to validate feature importance claims: The authors developed a significance testing procedure using three-cumulant chi-square approximation that works reliably even with small sample sizes, sparse data, and skewed distributions—common challenges in real-world datasets. This means organizations can now distinguish between genuinely important feature groups and those that appear important by chance, reducing the risk of making decisions based on spurious patterns.

Group Shapley distributes importance more fairly across features: Using Lorenz curves and Gini indices, the paper demonstrates that Group Shapley assigns feature importance more equitably compared to traditional individual Shapley values. This fairer allocation is important for responsible AI because it prevents over-reliance on a small number of features and provides a more balanced view of what drives model predictions.

Practical validation shows market variables dominate bond recovery predictions: In the real-world application analyzing 2,094 bond observations across 98 features, market-related variables emerged as the most influential group for predicting recovery rates. This finding demonstrates the method's practical utility and suggests that bond recovery is more sensitive to current market conditions than to static firm characteristics or macroeconomic factors.