Error, Noise, and Bias of Auditors’ Going Concern Opinions and the Role of Machine Learning
30 Pages Posted: 15 Dec 2021
Date Written: December 2021
This paper studies the error, bias, and noise of auditors’ Going Concern Opinion (GCO) and Machine Learning (ML) models in predicting firm default. We find that advanced ML models can significantly reduce the error, bias, and noise of default predictions compared to GCO, consistent with the theory in Kahneman, Sibony, & Sunstein (2021). Following Kahneman et al. (2021), we also explore the value of diversity in improving prediction quality. To that end, we construct four “artificial auditors” representing Big4 auditors. We find that the consensus from these artificial auditors can significantly reduce the prediction error compared to GCO. Our study adds to the accounting literature by examining the quality of GCO and the mechanism through which ML improves default prediction from the angle of prediction error, bias, and noise.
Keywords: Going Concern Opinion, Machine Learning, Prediction Error, Prediction Bias, Prediction Noise
JEL Classification: M41, M42
Suggested Citation: Suggested Citation