Detecting Earnings Management through Explainable Artificial Intelligence: A Machine Learning Approach Using Financial Statement Data
International Journal of Digital Accounting and Fintech Sustainability · 30.06.2026 · 3 (1) · 284-299
DOI: 10.70568/ijdafs.3.1.1 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Digital Accounting and Fintech Sustainability, online erschienen am 30.06.2026. Verfasst von Safwat Al Tal. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
This paper develops an explainable machine learning framework to classify the firm-year observations with unusually high discretionary accruals, a common proxy for the accrual-based earnings management, by using the financial statement data. The sample includes 7,800 firm-year observations of publicly listed non-financial firms from Compustat North America (via WRDS) for firms listed on the NYSE, Nasdaq and NYSE American over the fiscal years 2010-2023. The discretionary accruals were estimated by the Modified Jones Model, and they were used to build a proxy-based binary label according to the top quintile of the absolute discretionary accruals inside each industry-year (1,560 observations, 20.0%, classified as high-accrual). Four algorithms, which are the Logistic Regression, Random Forest, XGBoost and LightGBM, were trained on a leakage-screened set of financial ratios that excludes every component used in the label, and they were compared with the Beneish M-Score. The models were trained on the 2010-2018 data, tuned on 2019-2020, and evaluated out-of-sample on 2021-2023 (1,000 observations). The XGBoost achieves the highest ROC-AUC (0.748) and PR-AUC (0.421), which is higher than the Beneish M-Score (0.612), while the LightGBM achieves the highest F1-score (0.484) and balanced accuracy (0.653) at the 0.50 threshold, and the ROC-AUC advantage of the ML models is statistically significant (DeLong test, p = 0.004). Moreover, the SHAP analysis of the optimal model shows that the days sales uncollected, the gross profit margin and the debt-to-equity ratio are the leading variables, and an illustrative walkthrough shows how the SHAP output can be turned into a line-item diagnostic report. The findings indicate that the explainable tree-based models can improve the classification of the high-accrual firm-years compared to the Beneish M-Score, while keeping the interpretability that the auditors and the regulators need.
Zitieren
APA
Tal, S. A. (2026). Detecting Earnings Management through Explainable Artificial Intelligence: A Machine Learning Approach Using Financial Statement Data. International Journal of Digital Accounting and Fintech Sustainability, 3(1), 284-299. https://doi.org/10.70568/ijdafs.3.1.1
Vancouver
Tal SA. Detecting Earnings Management through Explainable Artificial Intelligence: A Machine Learning Approach Using Financial Statement Data. International Journal of Digital Accounting and Fintech Sustainability. 2026;3(1):284-299. doi:10.70568/ijdafs.3.1.1
BibTeX
@article{Tal2026,
author = {Tal, Safwat Al},
title = {Detecting Earnings Management through Explainable Artificial Intelligence: A Machine Learning Approach Using Financial Statement Data},
journal = {International Journal of Digital Accounting and Fintech Sustainability},
year = {2026},
volume = {3},
number = {1},
pages = {284--299},
doi = {10.70568/ijdafs.3.1.1}
}
EUVIMED