Predicting Accounting Fraud in U.S. Publicly Traded Firms: A Machine Learning Approach to Financial Misstatements

Sammanfattning

This study investigates whether machine learning models can predict accounting fraud in U.S. publicly traded firms, and whether the inclusion of predictorsrelated to intangible assets and cash flow items improves out-of-sample predictive performance. We apply a Random Under-Sampling Boosting (RUSBoost) algorithm to a sample of 255,052 firm-year observations covering theperiod 2000-2025, with 2,102 fraud observations identified through the SEC’sAccounting and Auditing Enforcement Releases (AAERs). To construct ourfraud sample, we leverage the recent advancements in large language models(LLMs) to extract violation periods from AAER documents. Our baselinespecification, referred to as the Baseline Model, uses 28 raw financial statement items as predictors and achieves an average AUC of 0.739 and an averageNDCG@1% of 0.048. The inclusion of predictors related to intangible assetssignificantly improves performance, raising the average AUC to 0.745 and theaverage NDCG@1% to 0.053, with goodwill emerging as the most importantfeature in the model. In contrast, the inclusion of cash flow items does notimprove predictive performance. We also quantify the serial fraud memorisation problem, showing that improper handling of serial fraud substantiallyinflates predictive performance. In addition, we demonstrate the importanceof controlling for seed-level variation when using flexible ensemble learningalgorithms.Our findings contribute to the accounting fraud detection literature bydemonstrating that balance sheet composition carries fraud-relevant information beyond what aggregated financial figures capture, that LLM-based extraction offers a viable and cost-efficient alternative to manual curation ofAAER samples and that methodological choices around serial fraud and seedvariation materially affect reported model performance.