Forward-Looking Financial Reporting Risk Screening in Saudi Listed Firms: Evidence from the Beneish M-Score and Explainable Machine Learning

Abstract

Most applications of the Beneish M-Score are contemporaneous: they screen financial statements after a reporting year has ended. This study asks whether the eight Beneish indicators measured in year t contain information about the following year's Beneish screening classification. The analysis uses a balanced panel of 66 Saudi listed non-financial firms from 2018 to 2025, yielding 528 annual financial statements and 462 M-Score observations for 2019-2025. Under the primary M > -1.78 rule, 66 of 462 firm-years (14.3%) generate a screening signal. On the untouched 2025 holdout, logistic regression achieves ROC-AUC = 0.718 and average precision = 0.371, while Random Forest achieves ROC-AUC = 0.684 and average precision = 0.406 and produces the higher precision and F1 at the pre-specified 0.50 decision-score threshold. A rolling-origin stability analysis across 2021-2025 uses the continuous current-year M-Score as the benchmark: pooled ROC-AUC is 0.537 for M_t, 0.661 for logistic regression, and 0.653 for Random Forest. Firm-cluster bootstrap inference indicates that both fitted models improve discrimination relative to M_t, whereas their difference is not statistically distinguishable. Unweighted refits produce similar ranking performance. SHAP identifies TATA as the leading Random Forest feature, followed by DEPI, AQI, and DSRI. The evidence supports explainable machine learning as a complementary prioritization tool rather than a calibrated fraud probability or a substitute for professional review.