Machine Learning for Financial-Statement Anomaly Detection in U.S. Small and Mid-Sized Enterprises: A Quantitative Study Using Gradient Boosting and SHAP

Abstract

Financial-statement anomalies can signal reporting irregularities, yet identifying them among small and mid-sized enterprises (SMEs) is challenging because labeled cases are uncommon and financial relationships may be nonlinear. This quantitative study examined the classification performance of gradient boosting for firm-year anomaly detection among U.S. SMEs and used SHAP to explain model predictions. The planned analytical framework compared Boost with logistic regression and random forest using financial-statement predictors. Evaluation included precision, recall, specificity, F1 score, ROC-AUC, PR-AUC, Brier score, and paired comparisons. Validation was designed to separate firms and reporting periods across model development and testing, while SHAP summaries and local explanations characterized predictor contributions. The numerical results in this draft are examples, not estimates calculated from an analyzed dataset. In the example, the sample contained 1,200 firm-year observations from 240 SMEs, including 96 labeled anomalous observations, or 8.0%. The test set contained 260 observations, with anomaly prevalence of 7.7%. Boost achieved ROC-AUC of 0.914 and PR-AUC of 0.611, compared with 0.831 and 0.412 for logistic regression, and 0.879 and 0.524 for random forest. Its example recall was 0.800, precision was 0.471, and Brier score was 0.066. Paired comparisons showed an ROC-AUC difference of 0.083 versus logistic regression (95% CI [0.018, 0.146], p = .014) and 0.035 versus random forest (95% CI [−0.024, 0.094], p = .230). Receivable’s growth, accrual intensity, and operating cash flow relative to earnings ranked among the leading SHAP predictors. These example findings demonstrate how discrimination, calibration, threshold performance, and model explanations can be reported together. Actual conclusions require analysis of the study data, verification of outcome labels, and replacement of every estimate with observed results. SHAP attributions describe model behavior and do not establish causality, misstatement, or fraud.