Detecting Accounting Fraud in Publicly Traded U.S. Firms Using a Machine Learning Approach

Abstract

We develop a state-of-the-art fraud prediction model using a machine learning approach. We demonstrate the value of combining domain knowledge and machine learning methods in model building. We select our model input based on existing accounting theories, but we differ from prior accounting research by using raw accounting numbers rather than financial ratios.We employ one of the most powerful machine learning methods, ensemble learning, rather than the commonly used method of logistic regression. To assess the performance of fraud prediction models, we introduce a new performance evaluation metric commonly used in ranking problems that is more appropriate for the fraud prediction task. Starting with an identical set of theory-motivated raw accounting numbers, we show that our new fraud prediction model outperforms two benchmark models by a large margin: the Dechow et al. logistic regression model based on financial ratios, and the Cecchini et al. support-vector-machine model with a financial kernel that maps raw accounting numbers into a broader set of ratios. JEL codes: C53; M41

What Are You Saying? Using topic to Detect Financial Misreporting

“…This evidence is consistent with Bao et al. [], who document that the AAER sample is more stable prior to 2005, due to the long time lag in SEC investigations and possible changes in the SEC's enforcement priorities after the 2008 financial crisis. We observe a similar post‐2005 decline in the AA sample.…”

“…We follow prior research and use the area under the receiver operating characteristics (ROC) curve to evaluate the out‐of‐sample classification performance of each detection model (Hobson, Mayew, and Venkatachalam [], Larcker and Zakolyukina [], Bao et al. []). The ROC curve is a two‐dimensional plot across different cutoff thresholds of the true positive rate (sensitivity) on the y ‐axis against the false positive rate (specificity) on the x ‐axis.…”

“…In addition, we follow Bao et al. [] and use the Normalized Discounted Cumulative Gain at the position k (NDCG@k) as an alternative measure of classification accuracy, where k is the top 1% or 99th percentile of the predicted probability scores. The NDCG@k measure uses a logarithmic function to weight the ranks of the predicted scores from a given model, such that the top prediction scores have higher values than lower prediction scores.…”

“…[], Bao et al. []), the rate of misreporting is very low, ranging from 1.34% to 1.65% across all years in the three samples and from 1.18% to 1.89% across the prediction years. The rate of misreporting in the AAER sample declines substantially after 2005.…”

“…[], Bao et al. []). One drawback of this approach is that financial misreporting can go undetected for multiple periods, because misreporting firms often manipulate performance metrics and accounting transactions to blend in better with their peers or the firm's own past performance (Lewis []).…”