New statistical test brings rigor to grouped AI explanations in finance

Machine learning models now sit at the heart of financial prediction, but their inner workings remain stubbornly opaque. A new study published in Humanities and Social Sciences Communications tackles one of the most practical barriers to trustworthy artificial intelligence in finance: how to decide, with statistical confidence, whether an entire category of predictors genuinely matters to a model’s output. The research, led by Jingyi Wang and Ying Chen of the National University of Singapore together with Paolo Giudici of the University of Pavia, develops a formal inferential framework for grouped feature attribution and demonstrates its value on a notoriously difficult problem, predicting how much money creditors recover when a bond defaults.

The starting point is Shapley value attribution, a technique borrowed from cooperative game theory that has become the de facto standard for explaining machine learning predictions. Shapley values allocate a model’s prediction across its inputs by simulating how much each feature contributes when combined with every possible subset of the other features. In principle, this gives a fair, mathematically grounded measure of importance. In practice, however, modern financial datasets contain dozens or even hundreds of features, and explaining a prediction one variable at a time produces a flood of granular numbers that can be too noisy to interpret reliably. Individual attributions can also be unstable when features are correlated, a pervasive condition in economic data where firm fundamentals, market prices, and macroeconomic indicators move together.

The authors’ solution is to organize predictors into economically meaningful groups and to ask a sharper question: is the combined attribution of an entire group statistically significant, or could it arise by chance? This reframing turns explanation from a descriptive exercise into a hypothesis test. Rather than simply ranking groups by their average attribution, the framework assesses whether group-level SHAP-style importance measures differ from what would be expected under a null hypothesis of no contribution. The procedure is built on a three-cumulant chi-square approximation, a technical device that corrects the distribution of the test statistic for the skewness and kurtosis that plague real-world attribution data. Standard approximations that rely on the first two moments alone can misstate significance levels in exactly the settings where explanation matters most.

The challenge is formidable. The authors designed the testing procedure for high-dimensional settings involving skewed distributions, sparse signals, dependence among features, and limited sample sizes, conditions that describe most empirical finance research. Simulation experiments show that the proposed test maintains accurate empirical size, meaning it does not falsely flag unimportant groups more often than its nominal significance level would allow. Perhaps more striking, the test often achieves higher power than competing alternatives such as the Wald test and the CQ test, meaning it is better at detecting genuine group importance when it exists. In statistical testing, there is an inherent tension between these two properties, and a method that holds the line on false positives while catching more true signals represents a meaningful advance.

The paper also contributes a tree-based grouped implementation designed for ensemble models, the family of algorithms that includes random forests and gradient boosting machines and that dominates applied prediction tasks. Ensemble methods are prized for accuracy but are among the hardest to explain, and extending grouped attribution to tree ensembles means the inferential framework can be deployed on the models practitioners actually use rather than on simplified stand-ins. This pairing of a rigorous test with a practical implementation is what elevates the work from methodological curiosity to usable tool.

To demonstrate the framework in action, the researchers turned to bond recovery rate prediction, a problem with direct consequences for investors, regulators, and the stability of the financial system. When a bond defaults, the recovery rate determines how much of the owed principal and interest creditors ultimately receive. Accurate recovery rate forecasts feed into bond pricing, credit risk models, and capital requirements, yet the determinants of recovery are contested. The study assembled a global dataset spanning 1996 to 2023, comprising 2094 observations and 98 features. The features were organized into 16 subgroups and five broader categories: bond characteristics, firm fundamentals, industry-specific factors, market-related variables, and macroeconomic indicators.

The empirical results deliver a clear verdict. Market-related variables and bond characteristics emerge as the most influential groups in predicting recovery rates. This finding carries weight precisely because it is backed by formal significance testing rather than by eyeballing attribution magnitudes. In a dataset with 98 features, many individual variables would show nonzero attributions simply by chance or through correlation with genuinely important factors. Group-level inference cuts through this ambiguity by pooling evidence across economically coherent categories, allowing the data to speak about whether market conditions as a whole, or bond terms as a whole, systematically shape what creditors recover. The result aligns with the intuition that recovery outcomes depend heavily on the state of markets at the time of default and on the specific terms of the defaulted instrument.

The implications extend well beyond credit markets. Grouped feature attribution is increasingly important in any high-dimensional application where predictors belong to scientifically meaningful categories, whether that means genes in biological pathways, sensors in engineering systems, or survey items in social research. The framework offers a statistically rigorous approach to global inference on grouped SHAP-style importance measures, addressing a gap that has limited the interpretability of machine learning in the social sciences. Researchers in humanities and social sciences communications, the journal’s own remit, often work with exactly the kind of data that strains individual-level explanations: many correlated variables, skewed distributions, and modest samples.

What makes the work notable for the broader conversation about artificial intelligence accountability is its insistence that explanations themselves be held to inferential standards. As machine learning models influence lending decisions, risk assessments, and policy, regulators and the public increasingly demand to know why a model produced its output. But an explanation that cannot distinguish signal from noise is a fragile foundation for accountability. By providing a test that maintains accurate empirical size under challenging conditions and often outperforms established alternatives, the study gives analysts a way to say not just which groups of factors a model appears to use, but which groups it uses with statistical confidence. In finance, where decisions ripple through markets and institutions, that distinction could prove decisive.

The research involved no external funding, and the authors declare no competing interests. All data are secondary and publicly accessible, and the study did not require ethical approval because it involved no human participants or animal subjects. The article is published open access, making the framework, its tree-based implementation, and the bond recovery application available to researchers, practitioners, and regulators who need explanations they can trust. As machine learning continues its advance into high-stakes domains, the ability to test whether entire categories of predictors matter, rather than merely observing that they seem to, may become as fundamental to responsible data science as the models themselves.

Subject of Research: Statistics and machine learning methodology with application to financial risk prediction

Article Title: Group Shapley values, with application to bond recovery rate prediction

Article References: Wang, J., Giudici, P., & Chen, Y. (2026). Group Shapley values, with application to bond recovery rate prediction. Humanities and Social Sciences Communications. https://doi.org/10.1057/s41599-026-08451-x

Image Credits: AI Generated

DOI: 10.1057/s41599-026-08451-x

Keywords: Shapley values, explainable AI, feature attribution, bond recovery rates, statistical inference, machine learning, credit risk, chi-square approximation, Group, Shapley, values, application