ACCT 660 Week 8 Discussion Board Forum 4 Replies Example

Discussion Board Forum 4 Replies: Responding to Classmates on Machine Learning in Fraud Detection

ACCT 660 Information Technology and Fraud, Discussion Board Forum 4 Replies

Thread by [Student Name]

The classmates and their threads are composites; the studies cited are real.

Reply to Tyler

Tyler, your thread predicted that machine learning will soon replace auditors' judgment in detecting financial statement fraud. I agree that the models are improving, but the research suggests they will change where auditors spend their judgment rather than remove it.

Perols (2011) compared several statistical and machine learning methods for predicting financial statement fraud under realistic assumptions about how rare fraud is and how costly it is to miss. Logistic regression and support vector machines performed well relative to more complex methods such as neural networks and stacking, which is a useful reminder that complexity does not guarantee accuracy. More recently, Bao et al. (2020) trained an ensemble learning model on raw financial data items from public companies and found that it outperformed earlier models, including a well-known ratio-based score. Even so, the best model flagged many firms for each one that turned out to have committed fraud. A model that narrows a thousand companies to fifty still leaves fifty for someone to judge.

That is where I think judgment remains essential. An auditor or examiner has to decide which flagged firms deserve investigation, what evidence to gather and how to explain a conclusion to a court or audit committee. My question for you: if a model flags a client but the audit team finds nothing after reasonable work, how should the auditor document that judgment? I ask because the documentation is where the auditor's reasoning becomes visible to reviewers and regulators, and a model's flag that is dismissed without a clear record could look like a missed warning if fraud surfaces later.

Reply to Hannah

Hannah, you proposed training machine learning models on past fraud cases, such as companies the SEC has charged, so that the models learn what fraud looks like. That is how most research models are built, and Dechow et al. (2011) used SEC enforcement releases in exactly this way to develop a misstatement score. I would add one limitation that matters a great deal.

The models can learn only from frauds that were caught. Dyck et al. (2024) used an unexpected event, the collapse of the audit firm Arthur Andersen, which forced its clients to change auditors and exposed more of their misconduct, to estimate how much fraud normally goes unseen. They estimated that only about one third of corporate frauds are detected and that a meaningful share of large public companies commit securities fraud in a given year. If that estimate is close, a model trained on detected cases learns the patterns of careless or unlucky fraudsters and may miss the careful ones who were never caught.

That does not make the models useless, but it suggests combining them with methods that do not depend on past cases, such as the anomaly tests in Nigrini's text. My question for you: how could an examiner check whether a model is missing a kind of fraud that has never been detected? One possibility is to test the model on cases uncovered later by whistleblowers or journalists, since those frauds were missed by the usual detection channels and may look different from the ones regulators found.

References

Bao, Y., Ke, B., Li, B., Yu, Y. J., & Zhang, J. (2020). Detecting accounting fraud in publicly traded U.S. firms using a machine learning approach. Journal of Accounting Research, 58(1), 199-235. https://doi.org/10.1111/1475-679X.12292

Dechow, P. M., Ge, W., Larson, C. R., & Sloan, R. G. (2011). Predicting material accounting misstatements. Contemporary Accounting Research, 28(1), 17-82. https://doi.org/10.1111/j.1911-3846.2010.01041.x

Dyck, A., Morse, A., & Zingales, L. (2024). How pervasive is corporate fraud? Review of Accounting Studies, 29(1), 736-769. https://doi.org/10.1007/s11142-022-09738-5

Perols, J. (2011). Financial statement fraud detection: An analysis of statistical and machine learning algorithms. Auditing: A Journal of Practice & Theory, 30(2), 19-50. https://doi.org/10.2308/ajpt-50009