Detection of fraudulent financial reports with machine learning techniques

This paper describes our efforts to apply various advanced supervised machine learning and natural language processing techniques, including Binomial Logistic Regression, Support Vector Machines, Neural Networks, Ensemble Techniques, and Latent Dirichlet Allocation (LDA), to the problem of detecting fraud in financial reporting documents available from the United States´ Security and Exchange Commission EDGAR database. Specifically, we apply LDA to a collection of type 10-K financial reports and to generate document-topic frequency matrix, and then submit these data to a series of advanced classification algorithms. We then apply evaluation metrics, such as Precision, Receiver Operating Characteristic Curve, and Area Under the Curve to evaluate the performance of each algorithm. We conclude that these methods show promise and suggest applying the approach to a larger set of input documents.