Going concern prediction – A horse race between traditional and regularization machine learning models

Abstract

Regularization machine learning (ML) methods have been increasingly applied in accounting research, offering new possibilities in predictive modeling. Their forte lies in the effective regularization methods for resolving the biggest concern of generalization, which is the risk of overfitting the training data. While these sophisticated methods are known to outperform traditional regression approaches in large and balanced datasets, this may not be the case when facing imbalanced and small datasets. Moreover, model validation is also challenging in such settings because traditional performance measures, such as prediction accuracy, may be misleading. We address this problem by comparing two traditional and five regularization-based methods in predicting going concern uncertainty (GCU) on the sample of listed companies in Croatia. We take caution when evaluating the models due to class-imbalanced problems and include different classification performance measures, as well as calibration of the models to account for their uncertainty. As expected, no model performs best across all evaluation criteria, but regularization methods are better calibrated. Given our results, we suggest that model selection should consider the results of the model calibration, a combination of different performance metrics, and the economic impact of the statistical performance of the model, if feasible.