Detecting Fraud in Financial Reports

Fraud in public companies has a large financial impact, and yet is only weakly detected by those who look for it, many incidents have been detected only when whistleblowers have come forward. We examine the problem of detecting fraud from the textual component of the quarterly and annual reports that public companies are required to file. Using an empirically derived set of words, we achieve prediction accuracy up to 88% on a per-report basis. Frauds rarely involve only a single quarter, so it is actually more useful to consider prediction performance on a per-incident basis. The truthfulness probability of our measure shows consistent decreases in the quarters leading up to a fraud, creating opportunities for proactive enforcement. We also compare the prediction performance of our word list with Pennebaker's deception model, and with a set of fixed lists suggested in the literature, only two of which have any predictive power.