Machine learning to detect fraud-indicating anomalies in corporate financial statements

Resumen

Financial statement fraud poses a threat to the transparency of financial information reported in the markets and to investor confidence. Given this scenario, the development and implementation of AI-based systems that incorporate ML techniques allow for the analysis of large volumes of financial data and the early detection of anomalies and patterns associated with fraudulent practices in financial reports. The CRISP-ML(Q) methodology is used, along with a database of 146,045 records from the SEC and COMPUSTAT, with a marked imbalance of fraud cases (0.00066%). Supervised (random forest, RF; logistic regression, LR) and unsupervised (isolation forest, IF; one-class SVM; local outlier factor, LOOF) models were used, along with feature selection, normalization, and oversampling with SMOTE. The results show that the RF model achieved a high overall accuracy of 98%, standing out for its ability to correctly classify most observations. Unsupervised models achieved a high level of recall in fraud detection (80%), demonstrating their effectiveness in identifying a greater number of fraudulent cases. It is concluded that effective fraud detection requires more robust approaches that combine traditional techniques with cost-sensitive algorithms, adaptive thresholds, and supervised models, strengthening anomaly identification and supporting financial risk auditing and management.