EXPLAINABLE MACHINE LEARNING EARLY-WARNING SYSTEM FOR CORPORATE GOVERNANCE FAILURE AND FINANCIAL RESTATEMENT DETECTION
Assaf Filfilan
Abstract: Corporate governance failures that end in financial restatement are costly, slow to surface and difficult for boards, auditors and regulators to anticipate. This study develops and evaluates an explainable machine learning early-warning system (EWS) that ranks listed firms by their risk of a material misstatement that is later the subject of a Securities and Exchange Commission (SEC) Accounting and Auditing Enforcement Release (AAER). Using a public benchmark panel of 108,399 US firm-years (15,320 firms, 1991-2008) with 837 misstatement firm-years, we train models only on information available two years before each test year and evaluate them out of sample over 2003-2008 (35,166 firm-years, 261 misstatements). A gradient-boosted tree model (XGBoost) built on 28 raw financial statement items achieves a mean out-of-sample AUC of 0.718 (pooled AUC 0.722, 95% CI 0.698-0.748), outperforming a logistic regression on the 14 financial ratios of Dechow et al. (2011) by 0.040 AUC points (95% CI 0.009-0.070; bootstrap p = .004). Translated into a four-tier EWS, the 20% of firms flagged Red, Amber or Watch contain 44.1% of all subsequently enforced misstatements, while the Green tier carries a misstatement rate 30% below the base rate. Shapley additive explanations (SHAP) show that the model relies on a small, stable set of signals (low property, plant and equipment intensity, high share price, recent equity issuance, large total and current assets), with year-to-year rank correlations of 0.84-0.98. The ratio-based logit remains stronger at the extreme top 1% of the ranking, which supports a layered design in which a flexible model screens broadly and a transparent model prioritises. The paper contributes a reproducible, explainable and temporally honest EWS template, and documents its limits: modest absolute precision, label incompleteness and the absence of board-level governance variables.
Keywords: Corporate governance failure, Financial restatement, Accounting fraud detection, Early-warning system, Explainable artificial intelligence, SHAP, Gradient boosting; AAER
DOI: 10.24874/PES08.03B.017
Recieved: 29.04.2026 Revised: 05.08.2026 Accepted: 17.08.2026
UDC:
Reads: 8 
