Shapley in Context: Explaining Financial Language with Domain Expertise

As large language models (LLMs) become increasingly integrated into high-stakes financial applications, the need for transparency and explainability grows. This paper addresses the challenge of interpreting LLM outputs on financial text, where general-purpose explanation methods often fail to incorporate domain-specific knowledge. The authors investigate whether the Shapley value—a game-theoretic approach to feature attribution—can provide explanations that align with established financial reasoning. They propose a set of domain-inspired axioms, such as monotonicity and diminishing marginal effects, and prove that the baseline Shapley value (BShap) satisfies these criteria, making it a robust tool for financial interpretability.

The study establishes that BShap is uniquely suited for financial text because it operates on discrete features, unlike methods like Integrated Gradients that require differentiable, continuous inputs. The authors demonstrate that BShap preserves key properties, including:

• Individual Monotonicity: Ensures that features with a clear positive or negative impact on financial sentiment receive attributions with the corresponding sign.
• Pairwise Monotonicity: Respects the relative importance of different features, such as prioritizing gross margin over net earnings in specific growth-oriented financial contexts.
• First-order Monotonic Dominance: Allows for meaningful comparisons of how the same feature contributes differently across distinct models or companies.

Through empirical experiments on Form 10-K risk disclosures for firms like Silicon Valley Bank and Wells Fargo, the authors show that BShap successfully decomposes aggregate risk scores into interpretable components. These attributions align with historical financial events, such as the regulatory challenges faced by Wells Fargo or the liquidity pressures preceding the collapse of Silicon Valley Bank.

This work provides a rigorous framework for model risk management in finance. By ensuring that LLM explanations are not just mathematically sound but also financially intuitive, the authors enable practitioners to use LLMs with greater confidence in regulated environments. The ability to decompose complex, unstructured narrative data into actionable risk attributions offers a significant advantage for investors, regulators, and risk managers who must navigate evolving market uncertainties.