Explainable Earnings-Quality Risk Screening Using Hybrid Machine Learning and Governance Signals: An Auditor-Oriented Framework for Indian Listed Firms

Abstract

Artificial intelligence is increasingly used to prioritize audit attention, yet high predictive accuracy alone is insufficient in assurance settings where reviewers must understand why a firm-year has been classified as risky. This study develops an Explainable Earnings-Quality Risk Screening framework (EQR-XAI) for Indian listed-firm auditing. The framework is intentionally distinct from conventional financial-misstatement classifiers: it predicts an earnings-quality risk state rather than asserting fraud, combines accounting-ratio, cash-flow, governance, related-party, and auditor-transition signals, and produces auditor-readable explanations through SHAP attribution and rule-based reason codes. The research design specifies a firm-year panel assembled from publicly available annual reports and exchange disclosures, with a reproducible synthetic evaluation panel used in this paper to demonstrate the complete analytical workflow without representing simulated values as observed corporate facts. Comparative models include logistic regression, random forest, and gradient boosting. The illustrative experiment shows the proposed hybrid model reaching ROC-AUC 0.91 and F1 0.84, with calibration error below the benchmark models. Accrual intensity, the cash-flow-to-profit gap, receivable growth, related-party intensity, and auditor change emerge as the leading explanation drivers. The study contributes an auditor-oriented architecture that separates predictive screening from the professional conclusion, embeds explanation quality and calibration into model evaluation, and maps model outputs to review procedures. The framework is suitable for future validation on verified Indian firm-year enforcement, restatement, and qualified-report outcomes.