Detecting Earnings Manipulation in Banks Using Deep Learning Techniques: An Empirical Study from Iraq

This study builds a reproducible detector of earnings manipulation in Iraqi banks using a bank-year panel from 2010 to 2024 sourced from audited annual reports, IFRS 9 credit risk notes, Iraq Stock Exchange disclosures, and Central Bank of Iraq publications. The feature set aligns with banking mechanics discretionary loan loss provisioning residuals scaled by lagged loans, a three-year smoothing index between changes in NPL and provisions, asset growth, fee share dynamics, and leverage changes. The label flags the top quintile of discretionary provisioning within each year to focus on relative deviations. Data are winsorized within year, standardized on the training sample, and split chronologically into training 2010–2021 and testing 2022–2024. Two classifiers are compared a class-weighted logistic regression and a class-weighted SVM. Evaluation uses ROC-AUC, PR-AUC, F1, accuracy, and Brier score, with thresholds tuned on validation folds and probabilities calibrated. Results show that the SVM delivers stronger ranking and better operating tradeoffs than the logistic baseline when inputs are standardized and the decision threshold targets screening objectives. Out-of-sample gains appear in ROC-AUC and PR-AUC with a lower Brier score. Confusion matrices confirm higher specificity and controlled false alarms at useful recall. SHAP analysis validates economic interpretability delta leverage, asset growth, and DLLP drive the score, followed by fee share changes and smoothing. The framework supports an audit workflow that screens bank-years, routes alerts to document-level review of allowance movements and write-offs, and updates models annually with rolling windows while preserving time integrity and comparability.