EXACT CONCEPT-LEVEL EXPLANATIONS FOR EARLY DETECTION OF FINANCIAL MISSTATEMENTS
Keywords:
Financial misstatement, Accounting fraud detection, Explainable artificial intelligence, Shapley values, Monotone constraints, Early warningAbstract
Early warnings of material financial misstatements are useful only when auditors and regulators can understand why a firm is flagged. Machine learning models trained on financial statement data rank misstatement risk well, but their post-hoc explanations refer to individual line items, change when the model is retrained and often contradict established accounting evidence. This study proposes AGX, an accounting-grounded explainable model for the early detection of misstatement risk. AGX derives double-entry articulation features from consecutive financial statements, organises all predictors into ten accounting concepts, restricts every boosted tree to a single concept, imposes the monotone relations documented in the fraud-prediction literature and averages balanced ensembles on the log-odds scale. We prove that the resulting risk score decomposes exactly into concept contributions that coincide with group Shapley values and never contradict the imposed accounting priors. On the public AAER benchmark of 146,045 U.S. firm-years from 1990 to 2014, evaluated with a temporal design that uses only labels available at deployment, AGX matches the ranking accuracy of leading ensemble methods and detects the first year of misstatement episodes significantly better than LightGBM and the Explainable Boosting Machine (onset-year AUC of 0.722 versus 0.698 and 0.692). It is also the most robust model to measurement noise and missing prior-year data, and its explanations are exact, faithful, the most stable across retraining and free of theory violations.References
[1] Karpoff J M, Koester A, Lee D S, et al. Proxies and databases in financial misconduct research. The Accounting Review, 2017, 92: 129-163. DOI: 10.2308/accr-51766.
[2] Dechow P M, Ge W, Larson C R, et al. Predicting material accounting misstatements. Contemporary Accounting Research, 2011, 28: 17-82.
[3] Beneish M D. The detection of earnings manipulation. Financial Analysts Journal, 1999, 55: 24-36.
[4] 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: 199-235. DOI: 10.1111/1475-679X.12292.
[5] Bertomeu J, Cheynel E, Floyd E, et al. Using machine learning to detect misstatements. Review of Accounting Studies, 2021, 26: 468-519. DOI: 10.1007/s11142-020-09563-8.
[6] 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: 221-245. DOI: 10.2308/accr-51562.
[7] Lundberg S M, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems 30, 2017: 4765-4774.
[8] 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: 56-67. DOI: 10.1038/s42256-019-0138-9.
[9] Aas K, Jullum M, Løland A. Explaining individual predictions when features are dependent: More accurate approximations to Shapley values. Artificial Intelligence, 2021, 298: 103502. DOI: 10.1016/j.artint.2021.103502.
[10] Beneish M D, Lee C M C, Nichols D C. Earnings manipulation and expected returns. Financial Analysts Journal, 2013, 69: 57-82.
[11] Cecchini M, Aytug H, Koehler G J, et al. Detecting management fraud in public companies. Management Science, 2010, 56: 1146-1160. DOI: 10.1287/mnsc.1100.1174.
[12] Perols J. Financial statement fraud detection: An analysis of statistical and machine learning algorithms. Auditing: A Journal of Practice & Theory, 2011, 30: 19-50.
[13] Abbasi A, Albrecht C, Vance A, et al. MetaFraud: A meta-learning framework for detecting financial fraud. MIS Quarterly, 2012, 36: 1293-1327.
[14] Kim Y J, Baik B, Cho S. Detecting financial misstatements with fraud intention using multi-class cost-sensitive learning. Expert Systems with Applications, 2016, 62: 32-43.
[15] Purda L, Skillicorn D. Accounting variables, deception, and a bag of words: Assessing the tools of fraud detection. Contemporary Accounting Research, 2015, 32: 1193-1223. DOI: 10.1111/1911-3846.12089.
[16] Hajek P, Henriques R. Mining corporate annual reports for intelligent detection of financial statement fraud – a comparative study of machine learning methods. Knowledge-Based Systems, 2017, 128: 139-152. DOI: 10.1016/j.knosys.2017.05.001.
[17] Brown N C, Crowley R M, Elliott W B. What are you saying? Using topic to detect financial misreporting. Journal of Accounting Research, 2020, 58: 237-291.
[18] Craja P, Kim A, Lessmann S. Deep learning for detecting financial statement fraud. Decision Support Systems, 2020, 139: 113421. DOI: 10.1016/j.dss.2020.113421.
[19] Amiram D, Bozanic Z, Rouen E. Financial statement errors: Evidence from the distributional properties of financial statement numbers. Review of Accounting Studies, 2015, 20: 1540-1593. DOI: 10.1007/s11142-015-9333-z.
[20] Lennox C, Lisowsky P, Pittman J. Tax aggressiveness and accounting fraud. Journal of Accounting Research, 2013, 51: 739-778. DOI: 10.1111/joar.12002.
[21] Achakzai M A K, Peng J. Detecting financial statement fraud using dynamic ensemble machine learning. International Review of Financial Analysis, 2023, 89: 102827. DOI: 10.1016/j.irfa.2023.102827.
[22] Zavitsanos E, Mavroeidis D, Bougiatiotis K, et al. Financial misstatement detection: A realistic evaluation. Proceedings of the 2nd ACM international conference on AI in finance (ICAIF’21), 2021: 34:1-34:9.
[23] Ribeiro M T, Singh S, Guestrin C. “Why should I trust you?”: Explaining the predictions of any classifier. Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (KDD’16), 2016: 1135-1144.
[24] Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019, 1: 206-215. DOI: 10.1038/s42256-019-0048-x.
[25] Lou Y, Caruana R, Gehrke J, et al. Accurate intelligible models with pairwise interactions. Proceedings of the 19th ACM SIGKDD international conference on knowledge discovery and data mining (KDD’13), 2013.
[26] Caruana R, Lou Y, Gehrke J, et al. Intelligible models for HealthCare: Predicting pneumonia risk and hospital 30-day readmission. Proceedings of the 21st ACM SIGKDD international conference on knowledge discovery and data mining (KDD’15), 2015.
[27] Nori H, Jenkins S, Koch P, et al. InterpretML: A unified framework for machine learning interpretability. arXiv preprint arXiv:190909223, 2019.
[28] Hribar P, Collins D W. Errors in estimating accruals: Implications for empirical research. Journal of Accounting Research, 2002, 40: 105-134. DOI: 10.1111/1475-679X.00041.
[29] Ohlson J A. Earnings, book values, and dividends in equity valuation. Contemporary Accounting Research, 1995, 11: 661-687.
[30] Hirshleifer D, Hou K, Teoh S H, et al. Do investors overvalue firms with bloated balance sheets? Journal of Accounting and Economics, 2004, 38: 297-331. DOI: 10.1016/j.jacceco.2004.10.002.
[31] Chen T, Guestrin C. XGBoost: A scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (KDD’16), 2016: 785-794.
[32] Ke G, Meng Q, Finley T, et al. LightGBM: A highly efficient gradient boosting decision tree. Advances in neural information processing systems 30, 2017: 3146-3154.
[33] Liu X-Y, Wu J, Zhou Z-H. Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 2009, 39: 539-550.
[34] Shapley L S. A value for n-person games. In: Kuhn H W, Tucker A W, eds. Contributions to the theory of games II. Princeton University Press, 1953: 307-317.
[35] Owen G. Values of games with a priori unions. In: Henn R, Moeschlin O, eds. Mathematical economics and game theory: Essays in honor of oskar morgenstern. Springer, 1977: 76-88.
[36] 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: 185-197.
[37] Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 2011, 12: 2825-2830.