Predicting fraud in MD&A sections using deep learning

Conventional data analytic techniques have been successfully applied to detecting fraud in the Management's Discussion and Analysis sections of company filings mandated by the SEC. Here, we investigate whether fraud detection can be improved by applying deep learning techniques. We build 18 deep learning models and compare their performance on a set of MD&A documents. The best-performing model achieved an accuracy of 91% and an F1-score of 77%, only slightly better than a conventional XGBoost predictor that achieved an accuracy of 91% and an F1-score of 73%. Of the deep learning models, the transformer, those incorporating attention mechanisms, and convolutional neural networks performed well; somewhat surprisingly, sequential models such as LSTMs did not.