Paper: arXiv 2407.08953

Authors: Dangxing Chen, Yuan Gao

Abstract

Over the past few decades, machine learning models have been extremely successful. As a result of axiomatic attribution methods, feature contributions have been explained more clearly and rigorously. There are, however, few studies that have examined domain knowledge in conjunction with the axioms. In this study, we examine asset pricing in finance, a field closely related to risk management. Consequently, when applying machine learning models, we must ensure that the attribution methods reflect the underlying risks accurately. In this work, we present and study several axioms derived from asset pricing domain knowledge. It is shown that while Shapley value and Integrated Gradients preserve most axioms, neither can satisfy all axioms. Using extensive analytical and empirical examples, we demonstrate how attribution methods can reflect risks and when they should not be used.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 6.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents heavy mathematical formalization with axioms, proofs, and stochastic calculus concepts (like Greeks and Black-Scholes), while also using empirical examples and demonstrating practical implications for risk management.

Research Flowchart

  flowchart TD
  A["Research Goal: Validate if attribution methods<br>account for risk in asset pricing"] --> B["Domain: Finance & Risk Management"]
  B --> C["Methodology: Derive Asset Pricing Axioms"]
  C --> D["Process: Test Shapley Values<br>vs Integrated Gradients"]
  D --> E["Data: Empirical & Analytical Examples"]
  E --> F{"Key Finding"}
  F -->|"Shapley Values"| G["Preserves most axioms<br>but fails specific risk axioms"]
  F -->|"Integrated Gradients"| H["Preserves most axioms<br>but fails specific risk axioms"]
  G & H --> I["Outcome: Attribution methods require<br>domain-specific adjustments for risk"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes several axioms for reflecting risk in asset pricing and provides a thorough analysis of Shapley value (BShap) and Integrated Gradients, outlining their advantages and disadvantages.

Source sentence
Our work has made the following contributions: • Several axioms are proposed based on different aspects for reflecting risk in asset pricing. • Our examples demonstrate the motivations and signifi-cance of these axioms. • We provide a thorough analysis of Shapley value and IG, outlining their advantages and disadvantages. II.

Assumes 6

  • The feature space consists of n features. model
    Source sentence
    PRELIMINARIES For problem setup, assume we have n features.
  • The class of functions F is the set of real analytic functions. model
    Source sentence
    We denote a class of functions f : Rn →R by F. We assume F to be the set of real analytic functions.
  • Baseline point x′ is less than or equal to explicand x component‑wise (x ≥ x′). model
    Source sentence
    Without loss of generality (WLOG), we assume x ≥x′.
  • The training domain X is convex for the Convex Geometry Axiom. data
    Source sentence
    We say a BAM preserves convex geometry (CG) if X is convex, x, x′ ∈X, then the calculation of A(x, x′, f) only requires f(X).
Show 2 more
  • Option pricing features are limited to underlying price, risk‑free rate, time to expiration, strike, and volatility. data
    Source sentence
    Models are based on five features, namely underlying stock (index) prices S, risk-free interest rates r, time to expiration τ, strike prices K, and volatility σ such that x = (S, r, τ, K, σ).
  • Data are split randomly into 75 % training and 25 % test sets. execution
    Source sentence
    We randomly split data into 75% training data and 25% test data.

Data

  • 2008 U.S. European call and put option data · Wharton Research Data Services options proprietary
    2008 · U.S. equity options
    Source sentence
    D. Data and Neural Networks Setup for Option Pricing We collected call and option data from Wharton Research Data Services in 2008. There are 253 trading days with 123969 records of option data in total.
  • VIX index (30‑day expected volatility of S&P 500) macro public
    Source sentence
    B. VIX VIX is the ticker symbol for the Chicago Board Options Exchange’s CBOE Volatility Index. The VIX we use is the 30-day expected volatility of the S&P 500 index, more precisely the square root of a 30-day expected realized variance of the index.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • BShap preserves average individual monotonicity, demand individual monotonicity, (reverse) diminishing marginal effect, increasing marginal effect, and first‑order monotonic dominance. theoretical
    Source sentence
    BShap preserves average individual monotonic-ity (AIM), demand individual monotonicity (DIM), (reverse) diminishing marginal effect ((R)DME), increasing marginal effect (IME), and first-order monotonic dominance (FMD).
  • Empirical results on option pricing during the 2008 crisis validate that both BShap and IG preserve risk patterns such as Delta sign, Gamma convexity, and Rho direction. other figure from elsewhere in the paper
    Source sentence
    The results validated the Theorem V.1, V.2, and -100 -50 0 50 100 150 200 250 Attributions K t r S Features 2.61 189.57 -46.03 -4.65 -80.68 Fig. 2: IG for Eq 13 -100 -50 0 50 100 150 200 250 Attributions K t r S Features 1.96 183.4 -41.52 -8.5 -74.93 Fig. 3: BShap for Eq 13 demonstrated the potential of both IG and BShap in preserving risk patterns.
  • Integrated Gradients cannot maintain demand monotonicity, i.e., attributions may not increase with a monotonic feature. theoretical
    Source sentence
    IG, however, cannot maintain demand monotonicity, as if the model is monotonic in a certain feature, then the attributions for that feature increase as its value increases; Shapley value may involve calculations outside of the training domain, which could present difficulties such as times of financial crisis.
  • Shapley value may involve calculations outside the training domain, posing difficulties during financial crises. theoretical
    Source sentence
    IG, however, cannot maintain demand monotonicity, as if the model is monotonic in a certain feature, then the attributions for that feature increase as its value increases; Shapley value may involve calculations outside of the training domain, which could present difficulties such as times of financial crisis.
  • BShap may not be the best option when its preservation properties are insufficient for a given situation. negative
    Source sentence
    Therefore, BShap may not be the best option in this situation. VI.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 2 papers in the archive

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (158)

  1. Statistical classification methods in consumer credit scoring: a review (1997)
  2. A credit scoring model for personal loans (1989)
  3. Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research (2015)
  4. Benchmarking state-of-the-art classification algorithms for credit scoring (2003)
  5. The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients (2009)
  6. Multiple classifier architectures and their application to credit risk assessment (2011)
  7. An interpretable model with globally consistent explanations for credit risk (2018) — link
  8. Boolean decision rules via column generation (2018) — link
and 150 more
  1. ViCE: visual counterfactual explanations for machine learning models (2020)
  2. Significance tests for neural networks (2020)
  3. Selectivenet: A deep neural network with an integrated reject option (2019)
  4. Selective classification for deep neural networks (2017) — link
  5. Recent developments in consumer credit risk assessment (2007)
  6. The MNIST database of handwritten digits (1999)
  7. Deep learning (2015)
  8. Learning multiple layers of features from tiny images (2009)
  9. Introduction to reinforcement learning (1998)
  10. Pattern Recognition Letters (2005)
  11. On Optimal Recognition Error and Reject Tradeoff (1970)
  12. SelectiveNet: A Deep neural network with an integrated reject option (2019)
  13. Abstaining classification when error costs are unequal and unknown (2018) — link
  14. On optimal recognition error and reject Tradeoff (1970)
  15. A method for improving classification reliability of multilayer perceptrons (1995)
  16. On optimal reject rules and ROC curves (2005)
  17. An optimal reject rule for binary classifiers (2000)
  18. Bootstrap methods for reject rules of Fisher LDA (2006)
  19. An empirical comparison of ideal and empirical ROC-based reject rules (2007)
  20. Class noise vs. attribute noise: A quantitative study of their impacts (2004)
  21. Theoretical and experimental analysis of a Two-Stage system for classification (2002)
  22. Improving the classification accuracy of noisy dataset by effective data preprocessing (2018)
  23. Spectrally-normalized margin bounds for neural networks (2017)
  24. Multi-objective optimization for self-adjusting weighted gradient in machine learning tasks (2015) — link
  25. UCI machine learning repository (2017)
  26. Deep learning (2016)
  27. Show, attend and tell: Neural image caption generation with visual attention (2016) — link
  28. Very deep convolutional networks for large-scale image recognition (2015)
  29. Going deeper with convolutions (2015)
  30. Deep residual learning for image recognition (2016)
  31. Deep Reinforcement Learning: A brief survey (2017)
  32. Training deep neural-networks using a noise adaptation layer (2017)
  33. Making deep neural networks robust to label noise: A loss correction approach (2017)
  34. Understanding deep learning requires rethinking generalization (2017)
  35. Building high-level features using large scale unsupervised learning (2012)
  36. Unsupervised visual representation learning by context prediction (2015)
  37. The curious robot: Learning visual representations via physical interactions (2016)
  38. Colorful image colorization (2016)
  39. Webly supervised learning of convolutional networks (2015)
  40. Learning From noisy large-scale datasets With minimal supervision (2017)
  41. On the foundations of noise-free selective classification (2010)
  42. Selective classification for deep neural networks (2017) — link
  43. Support vector machines with embedded reject option (2002)
  44. The nearest neighbor classification rule with a reject option (1970)
  45. Boosting with abstention (2016)
  46. Reject option with multiple thresholds (2000)
  47. Neural networks for pattern recognition (1995)
  48. The multilayer perceptron as an approximation to a Bayes optimal discriminant function (1990)
  49. Methods for interpreting and understanding deep neural networks (2018)
  50. Sensitivity of feedforward neural networks to Weight Errors (1990)
  51. Sensitivity based Neural Networks Explanations (2018) — link
  52. Intelligible Models for Health-Care: Predicting Pneumonia Risk and Hospital 30-day Readmission (2015)
  53. Good Explanation for Algorithmic Transparency (2019)
  54. Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI (2019) — link
  55. Learning Algorithm in Two-Stage Selective Prediction (2022)
  56. Advances in financial machine learning (2018)
  57. Approximation by superpositions of a sigmoidal function (1989)
  58. Approximation capabilities of multilayer feedforward networks (1991)
  59. Neural networks: a comprehensive foundation by Simon Haykin, Macmillan, 1994, ISBN 0-02-352781-7. (1999)
  60. Fundamentals of artificial neural networks (1995)
  61. Neural networks for optimal approximation of smooth and analytic functions (1996)
  62. The eu general data protection regulation (gdpr) (2017)
  63. Machine learning for credit scoring: Improving logistic regression with non-linear decision-tree effects (2022)
  64. Credit granting: A comparative analysis of classification procedures (1987)
  65. Construction of a k-nearest-neighbour credit-scoring system (1997)
  66. Credit scoring using neural and evolutionary techniques (2000)
  67. Subagging for credit scoring models (2010)
  68. Using Random Forest for credit risk models (2018)
  69. Certified monotonic neural networks (2020)
  70. Sparse convolutional neural networks (2015)
  71. Deep sparse rectifier neural networks (2011)
  72. Training sparse neural networks (2017)
  73. A Biometrics invited paper. The analysis and selection of variables in linear regression (1976)
  74. Neural additive models: Interpretable machine learning with neural nets (2021)
  75. Enhancing explainability of neural networks through architecture constraints (2020)
  76. GAMI-Net: An explainable neural network based on generalized additive models with structured interactions (2021)
  77. Unwrapping the black box of deep ReLU networks: interpretability, diagnostics, and simplification (2020) — link
  78. A unified approach to interpreting model predictions (2017)
  79. " Why should i trust you?" Explaining the predictions of any classifier (2016)
  80. Fast and flexible monotonic functions with ensembles of lattices (2016)
  81. Deep lattice networks and partial monotonic functions (2017)
  82. Deep residual learning for image recognition (2016)
  83. Language models are unsupervised multitask learners (2019)
  84. Designing Inherently Interpretable Machine Learning Models (2021) — link
  85. Pitfalls to avoid when interpreting machine learning models (2020)
  86. Problems with Shapley-value-based explanations as feature importance measures (2020)
  87. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead (2019)
  88. Fooling lime and shap: Adversarial attacks on post hoc explanation methods (2020)
  89. Re: Request for Information and Comment on Financial Institutions’ Use of Artificial Intelligence, including Machine Learning (Docket No. CFPB-2021-0004) (2021)
  90. Seven challenges for harmonizing explainability requirements (2021) — link
  91. AI regulation is coming (2021) — [link](https://hbr.org/2021/09/ai-regulation- is- coming.)
  92. Model risk management (2021) — link
  93. Fair lending (2010) — link
  94. Multicriteria interpretability driven deep learning (2022)
  95. The many Shapley values for model explanation (2020)
  96. A rigorous study of integrated gradients method and extensions to internal neuron attributions (2022)
  97. Axiomatic attribution for deep networks (2017)
  98. Shape constraints for set functions (2019)
  99. Monotonic Neural Additive Models: Pursuing Regulated Machine Learning Models for Credit Scoring (2022)
  100. Three methods to share joint costs or surplus (1999)
  101. Anchors: High-precision model-agnostic explanations (2018)
  102. Physics-informed machine learning (2021)
  103. Hamiltonian neural networks (2019)
  104. Integrated Gradients is a Nonlinear Generalization of the Industry Standard Approach to Variable Attribution for Credit Risk Models (2022)
  105. Explainable machine learning in credit risk management (2021)
  106. Multidimensional shape constraints (2020)
  107. Applications of Integrated Gradients in Credit Risk Modeling (2022)
  108. Supermodularity and complementarity in economics: An elementary survey (2005)
  109. Explaining explanations: Axiomatic feature interactions for deep networks (2021)
  110. Review of the Federal Reserve’s Supervision and Regulation of Silicon Valley Bank (2023)
  111. Some theoretical problems suggested by the movements of interest rates, bond yields and stock prices in the United States since 1856 (1938)
  112. Fixed income analytics: state-of-the-art debt analysis and valuation modeling (1991)
  113. The pricing of options and corporate liabilities (1973)
  114. Stochastic calculus for finance II: Continuous-time models (2004)
  115. Diminishing returns shape constraints for interpretability and regularization (2018)
  116. Shape constrained additive models (2015)
  117. Studies of stock market volatility changes (1976)
  118. Notes on the n-person game—ii: The value of an n-person game.(1951) (1951)
  119. The shapley taylor interaction index (2020)
  120. Multilinear extensions of games (1972)
  121. An axiomatic approach to the concept of interaction among players in cooperative games (1999)
  122. A value for n-person games (1953)
  123. Values of non-atomic games (2015)
  124. How to address monotonicity for model risk management? — → 2305.00799
  125. Efficient capital markets: A review of theory and empirical work (1970)
  126. Option volatility & pricing: advanced trading strategies and techniques (1994)
  127. Derivatives on market volatility (1993)
  128. A closed-form solution for options with stochastic volatility with applications to bond and currency options (1993)
  129. A parametric nonlinear model of term structure dynamics (1999)
  130. The dynamics of stochastic volatility: evidence from underlying and options markets (2003)
  131. An equilibrium characterization of the term structure (1977)
  132. Volatility options in rough volatility models (2020)
  133. Deep learning volatility: a deep neural network perspective on pricing and calibration in (rough) volatility models (2021)
  134. Option return predictability with machine learning and big data (2023)
  135. Empirical asset pricing via machine learning (2020)
  136. Deep learning in asset pricing (2023)
  137. The leverage effect puzzle: Disentangling sources of bias at high frequency (2013)
  138. Theory of rational option pricing (1973)
  139. Can I Trust the Explanations? Investigating Explainable Machine Learning Methods for Monotonic Models — → 2309.13246
  140. Four Axiomatic Characterizations of the Integrated Gradients Attribution Method (2023) — link
  141. Handbook of the Shapley value (2019)
  142. A unifying framework to the analysis of interaction methods using synergy functions (2023)
  143. Interpretability in deep learning for finance: a case study for the Heston model (2021) — link
  144. Inferential theory for factor models of large dimensions (2003)
  145. Determining the number of factors in approximate factor models (2002)
  146. Using stocks or portfolios in tests of factor models (2020)
  147. Financial machine learning (2023)
  148. Options, futures, and other derivatives (2016)
  149. Incorporating Functional Knowledge in Neural Networks. (2009)
  150. The Shapley value decomposition of optimal portfolios (2021)

2 of 158 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.