EXPLAINABLE IDENTIFICATION OF TAX-RELATED FINANCIAL REPORTING ANOMALIES IN U.S. PUBLIC COMPANIES: CROSS-FITTED TAX-CONSISTENCY RESIDUALS AND OWEN-VALUE SHAPLEY PROPAGATION
Keywords:
Explainable artificial intelligence, Financial misstatement detection, Income tax accounting, Conditional anomaly detection, Shapley values, Owen valueAbstract
Income-tax accounts are reported under two regimes, financial accounting and the tax code, so their consistency with the rest of a firm’s financial statements is a natural check on reporting quality. Existing machine-learning detectors of financial misreporting treat tax items as generic inputs, and their explanations rarely reach the line items that an examiner can verify. We propose TRACE, an explainable framework for identifying tax-related financial reporting anomalies in U.S. public companies. First, for every fiscal year TRACE estimates label-free peer expectations of income-tax expense, taxes payable and their changes with cross-fitted conditional median and dispersion models, so that the resulting tax-consistency residuals cannot absorb the anomaly they measure. Second, it embeds these residuals in an imbalance-aware gradient-boosting ensemble that combines dollar-scale and within-year peer-rank representations of 28 Compustat items and is selected on validation years only. Third, it explains each risk score with Propagated Shapley Explanations, which chain exact single-baseline TreeSHAP, an exact Owen-value allocation over the reported, expected and dispersion components of each residual, and guarded propagation to 33 line items, with a proven completeness guarantee. On 146,045 U.S. firm-years labelled with SEC Accounting and Auditing Enforcement Releases and evaluated strictly out of time (2003–2008), TRACE achieves the highest NDCG at the 1% screening cut-off (0.0330), an improvement of 99% over RUSBoost and 89% over LightGBM (bootstrap p = 0.019 and 0.010) at comparable AUC. Its label-free tax-consistency score alone outperforms four generic outlier detectors. The explanations are exactly complete, agree closely with KernelSHAP (Spearman correlation 0.98) at 2.7 times lower cost, and show that detected misstatements most often combine tax expense growing faster than peers’ with taxes payable below peer expectations.References
[1] Dechow P M, Ge W, Larson C R, et al. Predicting material accounting misstatements. Contemporary Accounting Research, 2011, 28(1): 17-82.
[2] Bao Y, Ke B, Li B, et al. Detecting accounting fraud in publicly traded U.S. firms using a machine learning approach. Journal of Accounting Research, 2020, 58(1): 199-235.
[3] Bertomeu J, Cheynel E, Floyd E, et al. Using machine learning to detect misstatements. Review of Accounting Studies, 2021, 26(2): 468-519.
[4] Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019, 1(5): 206-215.
[5] Bertomeu J. Machine learning improves accounting: discussion, implementation and research opportunities. Review of Accounting Studies, 2020, 25(3): 1135-1155.
[6] Hanlon M. The persistence and pricing of earnings, accruals, and cash flows when firms have large book-tax differences. The Accounting Review, 2005, 80(1): 137-166.
[7] Blaylock B, Shevlin T, Wilson R J. Tax avoidance, large positive temporary book-tax differences, and earnings persistence. The Accounting Review, 2012, 87(1): 91-120.
[8] Graham J R, Raedy J S, Shackelford D A. Research in accounting for income taxes. Journal of Accounting and Economics, 2012, 53(1-2): 412-434.
[9] Hanlon M, Heitzman S. A review of tax research. Journal of Accounting and Economics, 2010, 50(2-3): 127-178.
[10] Frank M M, Lynch L J, Rego S O. Tax reporting aggressiveness and its relation to aggressive financial reporting. The Accounting Review, 2009, 84(2): 467-496.
[11] Lennox C, Lisowsky P, Pittman J. Tax aggressiveness and accounting fraud. Journal of Accounting Research, 2013, 51(4): 739-778.
[12] Erickson M, Hanlon M, Maydew E L. How much will firms pay for earnings that do not exist? Evidence of taxes paid on allegedly fraudulent earnings. The Accounting Review, 2004, 79(2): 387-408.
[13] Ribeiro M T, Singh S, Guestrin C. "Why should I trust you?": Explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016: 1135-1144.
[14] Lundberg S M, Lee S-I. A unified approach to interpreting model predictions. In: Advances in Neural Information Processing Systems 30, 2017: 4765-4774.
[15] Shrikumar A, Greenside P, Kundaje A. Learning important features through propagating activation differences. In: Proceedings of the 34th International Conference on Machine Learning, PMLR 70, 2017: 3145-3153.
[16] Chen H, Lundberg S M, Lee S-I. Explaining a series of models by propagating Shapley values. Nature Communications, 2022, 13: 4512.
[17] Owen G. Values of games with a priori unions. In: Henn R, Moeschlin O, eds. Essays in Mathematical Economics and Game Theory. Berlin: Springer, 1977: 76-88.
[18] Seiffert C, Khoshgoftaar T M, Van Hulse J, et al. RUSBoost: A hybrid approach to alleviating class imbalance. IEEE Transactions on Systems, Man, and Cybernetics – Part A: Systems and Humans, 2010, 40(1): 185-197.
[19] Walker S. Critique of an article on machine learning in the detection of accounting fraud. Econ Journal Watch, 2021, 18(1): 61-70.
[20] Bao Y, Ke B, Li B, et al. A response to "Critique of an article on machine learning in the detection of accounting fraud". Econ Journal Watch, 2021, 18(1): 71-78.
[21] Perols J L, Bowen R M, Zimmermann C, et al. Finding needles in a haystack: Using data analytics to improve fraud prediction. The Accounting Review, 2017, 92(2): 221-245.
[22] Grinsztajn L, Oyallon E, Varoquaux G. Why do tree-based models still outperform deep learning on typical tabular data? In: Advances in Neural Information Processing Systems 35 (Datasets and Benchmarks Track), 2022.
[23] Ke G, Meng Q, Finley T, et al. LightGBM: A highly efficient gradient boosting decision tree. In: Advances in Neural Information Processing Systems 30, 2017: 3146-3154.
[24] Chen T, Guestrin C. XGBoost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016: 785-794.
[25] Liu F T, Ting K M, Zhou Z-H. Isolation-based anomaly detection. ACM Transactions on Knowledge Discovery from Data, 2012, 6(1): 3.
[26] Goldstein M, Dengel A. Histogram-based outlier score (HBOS): A fast unsupervised anomaly detection algorithm. In: KI-2012: Poster and Demo Track, 2012: 59-63.
[27] Li Z, Zhao Y, Botta N, et al. COPOD: Copula-based outlier detection. In: 2020 IEEE International Conference on Data Mining (ICDM), 2020: 1118-1123.
[28] Li Z, Zhao Y, Hu X, et al. ECOD: Unsupervised outlier detection using empirical cumulative distribution functions. IEEE Transactions on Knowledge and Data Engineering, 2022.
[29] Zhao Y, Nasrullah Z, Li Z. PyOD: A Python toolbox for scalable outlier detection. Journal of Machine Learning Research, 2019, 20(96): 1-7.
[30] Paulheim H, Meusel R. A decomposition of the outlier detection problem into a set of supervised learning problems. Machine Learning, 2015, 100(2-3): 509-531.
[31] Chernozhukov V, Chetverikov D, Demirer M, et al. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 2018, 21(1): C1-C68.
[32] Lundberg S M, Erion G, Chen H, et al. From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence, 2020, 2(1): 56-67.
[33] Sundararajan M, Najmi A. The many Shapley values for model explanation. In: Proceedings of the 37th International Conference on Machine Learning, PMLR 119, 2020: 9269-9278.
[34] Petsiuk V, Das A, Saenko K. RISE: Randomized input sampling for explanation of black-box models. In: Proceedings of the British Machine Vision Conference (BMVC), 2018.