Paper: arXiv 2506.06653

Authors: Dangxing Chen

Abstract

In recent years, machine learning models have achieved great success at the expense of highly complex black-box structures. By using axiomatic attribution methods, we can fairly allocate the contributions of each feature, thus allowing us to interpret the model predictions. In high-risk sectors such as finance, risk is just as important as mean predictions. Throughout this work, we address the following risk attribution problem: how to fairly allocate the risk given a model with data? We demonstrate with analysis and empirical examples that risk can be well allocated by extending the Shapley value framework.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 3.0/10
  • Quadrant: Lab Rats — theoretically deep, empirically untested

Why this score: The paper heavily relies on advanced mathematical frameworks like Shapley value axioms, characteristic functions, and theoretical proofs, resulting in high mathematical complexity. However, the empirical section is described generically with ‘analysis and empirical examples’ without specifying backtests, datasets, or implementation details, indicating low empirical rigor.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Risk Attribution Problem: How to fairly allocate risk<br>given a model with data?"] --> B["Methodology<br>Extend Shapley Value Framework<br>(Axiomatic Attribution)"]
  B --> C["Input Data<br>Financial Model Predictions<br>Feature Datasets"]
  C --> D["Computational Process<br>Iterative Feature Combinations<br>Calculate Marginal Risk Contributions"]
  D --> E["Aggregation<br>Sum contributions via Shapley Formula<br>Ensure Efficiency & Symmetry"]
  E --> F["Key Outcomes<br>1. Fair Risk Allocation per Feature<br>2. Interpretable Risk Sensitivity<br>3. Validated on Financial Models"]
  F --> G["Conclusion<br>Risk accurately attributed<br>enhancing model interpretability in finance"]

Cited References (160)

  1. GRASP with path-relinking for the three-index assignment problem (2000)
  2. Probability distribution of solution time in GRASP: an experimental investigation (2003)
  3. Parallel GRASP with path-relinking for job shop scheduling (1984)
  4. Title of a Book in Another Language (Quoted in that Language) (1988)
  5. Multi-start and strategic oscillation methods – principles to exploit adaptive memory (2000)
  6. Bubbles in Japan’s stock markets: a macroeconomic analysis (1995)
  7. The resurgent Japanese economy and a Japan–United States free trade agreement (1997)
  8. Efficient algorithms for layer assignment problems (1996)
and 152 more
  1. Gravitation (1973)
  2. Parallel GRASP with path-relinking for job shop scheduling (1983)
  3. Title of Book (2002)
  4. Statistical classification methods in consumer credit scoring: a review (1997)
  5. A credit scoring model for personal loans (1989)
  6. Benchmarking state-of-the-art classification algorithms for credit scoring: An update of research (2015)
  7. Benchmarking state-of-the-art classification algorithms for credit scoring (2003)
  8. The comparisons of data mining techniques for the predictive accuracy of probability of default of credit card clients (2009)
  9. Multiple classifier architectures and their application to credit risk assessment (2011)
  10. An interpretable model with globally consistent explanations for credit risk (2018) — link
  11. Boolean decision rules via column generation (2018) — link
  12. ViCE: visual counterfactual explanations for machine learning models (2020)
  13. Significance tests for neural networks (2020)
  14. Selectivenet: A deep neural network with an integrated reject option (2019)
  15. Selective classification for deep neural networks (2017) — link
  16. Recent developments in consumer credit risk assessment (2007)
  17. The MNIST database of handwritten digits (1999)
  18. Deep learning (2015)
  19. Learning multiple layers of features from tiny images (2009)
  20. Introduction to reinforcement learning (1998)
  21. Pattern Recognition Letters (2005)
  22. On Optimal Recognition Error and Reject Tradeoff (1970)
  23. SelectiveNet: A Deep neural network with an integrated reject option (2019)
  24. Abstaining classification when error costs are unequal and unknown (2018) — link
  25. On optimal recognition error and reject Tradeoff (1970)
  26. A method for improving classification reliability of multilayer perceptrons (1995)
  27. On optimal reject rules and ROC curves (2005)
  28. An optimal reject rule for binary classifiers (2000)
  29. Bootstrap methods for reject rules of Fisher LDA (2006)
  30. An empirical comparison of ideal and empirical ROC-based reject rules (2007)
  31. Class noise vs. attribute noise: A quantitative study of their impacts (2004)
  32. Theoretical and experimental analysis of a Two-Stage system for classification (2002)
  33. Improving the classification accuracy of noisy dataset by effective data preprocessing (2018)
  34. Spectrally-normalized margin bounds for neural networks (2017)
  35. Multi-objective optimization for self-adjusting weighted gradient in machine learning tasks (2015) — link
  36. UCI machine learning repository (2017)
  37. Deep learning (2016)
  38. Show, attend and tell: Neural image caption generation with visual attention (2016) — link
  39. Very deep convolutional networks for large-scale image recognition (2015)
  40. Going deeper with convolutions (2015)
  41. Deep residual learning for image recognition (2016)
  42. Deep Reinforcement Learning: A brief survey (2017)
  43. Training deep neural-networks using a noise adaptation layer (2017)
  44. Making deep neural networks robust to label noise: A loss correction approach (2017)
  45. Understanding deep learning requires rethinking generalization (2017)
  46. Building high-level features using large scale unsupervised learning (2012)
  47. Unsupervised visual representation learning by context prediction (2015)
  48. The curious robot: Learning visual representations via physical interactions (2016)
  49. Colorful image colorization (2016)
  50. Webly supervised learning of convolutional networks (2015)
  51. Learning From noisy large-scale datasets With minimal supervision (2017)
  52. On the foundations of noise-free selective classification (2010)
  53. Selective classification for deep neural networks (2017) — link
  54. Support vector machines with embedded reject option (2002)
  55. The nearest neighbor classification rule with a reject option (1970)
  56. Boosting with abstention (2016)
  57. Reject option with multiple thresholds (2000)
  58. Neural networks for pattern recognition (1995)
  59. The multilayer perceptron as an approximation to a Bayes optimal discriminant function (1990)
  60. Methods for interpreting and understanding deep neural networks (2018)
  61. Sensitivity of feedforward neural networks to Weight Errors (1990)
  62. Sensitivity based Neural Networks Explanations (2018) — link
  63. Intelligible Models for Health-Care: Predicting Pneumonia Risk and Hospital 30-day Readmission (2015)
  64. Good Explanation for Algorithmic Transparency (2019)
  65. Explainable Artificial Intelligence (XAI): concepts, taxonomies, opportunities and challenges toward responsible AI (2019) — link
  66. Learning Algorithm in Two-Stage Selective Prediction (2022)
  67. Advances in financial machine learning (2018)
  68. Approximation by superpositions of a sigmoidal function (1989)
  69. Approximation capabilities of multilayer feedforward networks (1991)
  70. Neural networks: a comprehensive foundation by Simon Haykin, Macmillan, 1994, ISBN 0-02-352781-7. (1999)
  71. Fundamentals of artificial neural networks (1995)
  72. Neural networks for optimal approximation of smooth and analytic functions (1996)
  73. The eu general data protection regulation (gdpr) (2017)
  74. Machine learning for credit scoring: Improving logistic regression with non-linear decision-tree effects (2022)
  75. Credit granting: A comparative analysis of classification procedures (1987)
  76. Construction of a k-nearest-neighbour credit-scoring system (1997)
  77. Credit scoring using neural and evolutionary techniques (2000)
  78. Subagging for credit scoring models (2010)
  79. Using Random Forest for credit risk models (2018)
  80. Certified monotonic neural networks (2020)
  81. Sparse convolutional neural networks (2015)
  82. Deep sparse rectifier neural networks (2011)
  83. Training sparse neural networks (2017)
  84. A Biometrics invited paper. The analysis and selection of variables in linear regression (1976)
  85. Neural additive models: Interpretable machine learning with neural nets (2021)
  86. Enhancing explainability of neural networks through architecture constraints (2020)
  87. GAMI-Net: An explainable neural network based on generalized additive models with structured interactions (2021)
  88. Unwrapping the black box of deep ReLU networks: interpretability, diagnostics, and simplification (2020) — link
  89. A unified approach to interpreting model predictions (2017)
  90. " Why should i trust you?" Explaining the predictions of any classifier (2016)
  91. Fast and flexible monotonic functions with ensembles of lattices (2016)
  92. Deep lattice networks and partial monotonic functions (2017)
  93. Deep residual learning for image recognition (2016)
  94. Language models are unsupervised multitask learners (2019)
  95. Designing Inherently Interpretable Machine Learning Models (2021) — link
  96. Pitfalls to avoid when interpreting machine learning models (2020)
  97. Problems with Shapley-value-based explanations as feature importance measures (2020)
  98. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead (2019)
  99. Fooling lime and shap: Adversarial attacks on post hoc explanation methods (2020)
  100. Re: Request for Information and Comment on Financial Institutions’ Use of Artificial Intelligence, including Machine Learning (Docket No. CFPB-2021-0004) (2021)
  101. Seven challenges for harmonizing explainability requirements (2021) — link
  102. AI regulation is coming (2021) — [link](https://hbr.org/2021/09/ai-regulation- is- coming.)
  103. Model risk management (2021) — link
  104. Fair lending (2010) — link
  105. Multicriteria interpretability driven deep learning (2022)
  106. The many Shapley values for model explanation (2020)
  107. A rigorous study of integrated gradients method and extensions to internal neuron attributions (2022)
  108. Axiomatic attribution for deep networks (2017)
  109. Shape constraints for set functions (2019)
  110. Monotonic Neural Additive Models: Pursuing Regulated Machine Learning Models for Credit Scoring (2022)
  111. Three methods to share joint costs or surplus (1999)
  112. Anchors: High-precision model-agnostic explanations (2018)
  113. Physics-informed machine learning (2021)
  114. Hamiltonian neural networks (2019)
  115. How to address monotonicity for model risk management? (2023)
  116. Integrated Gradients is a Nonlinear Generalization of the Industry Standard Approach to Variable Attribution for Credit Risk Models (2022)
  117. Explainable machine learning in credit risk management (2021)
  118. Multidimensional shape constraints (2020)
  119. Applications of Integrated Gradients in Credit Risk Modeling (2022)
  120. Four Axiomatic Characterizations of the Integrated Gradients Attribution Method (2023) — link
  121. A value for n-person games (1953)
  122. A unifying framework to the analysis of interaction methods using synergy functions (2023)
  123. Producer incentives in cost allocation (1985)
  124. Coherent measures of risk (1999)
  125. Coherent risk measures on general probability spaces (2002)
  126. Generalized deviations in risk analysis (2006)
  127. Maximum drawdown (2004)
  128. Conditional value-at-risk for general loss distributions (2002)
  129. Inferential theory for factor models of large dimensions (2003)
  130. Determining the number of factors in approximate factor models (2002)
  131. Using stocks or portfolios in tests of factor models (2020)
  132. Robust portfolio optimization: a categorized bibliographic review (2020)
  133. Drawdown beta and portfolio optimization (2022)
  134. Alphabet’s Stock Tumbles as Cloud Slowdown Outweighs Strong Earnings (2023) — link
  135. Tesla Stock Plummets After Elon Musk Announces the Company’s Tech Boss Is Leaving (2019) — link
  136. Optimization of conditional value-at-risk (2000)
  137. Risk attribution using the Shapley value: Methodology and policy applications (2016)
  138. The Shapley value decomposition of optimal portfolios (2021)
  139. Portfolio risk allocation through Shapley value (2023)
  140. Theory of rational option pricing (1973)
  141. A closed-form solution for options with stochastic volatility with applications to bond and currency options (1993)
  142. The pricing of options and corporate liabilities (1973)
  143. Stochastic calculus for finance II: Continuous-time models (2004)
  144. Derivatives on market volatility (1993)
  145. Efficient capital markets: A review of theory and empirical work (1970)
  146. Risk and asset allocation (2005)
  147. Incorporating Functional Knowledge in Neural Networks. (2009)
  148. Attribution Methods in Asset Pricing: Do They Account for Risk? (2024) — link
  149. Sign symmetry vs symmetry: Young’s characterization of the Shapley value revisited (2018)
  150. Markowitz’s" portfolio selection": A fifty-year retrospective (2002)
  151. Variance allocation and Shapley value (2018)
  152. The CBOE volatility index-VIX (2009)

0 of 160 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2518 linked papers.