This study aims to examine how to improve the assessment and prediction of corporate going-concern capability, which remains a critical issue in auditing theory and practice. In particular, it addresses the challenge of identifying firms with significant going-concern difficulties among enterprises classified as high-risk.
The study develops an integrated social network analysis and machine learning (SNAML) framework to predict going-concern status. Textual analysis is conducted on auditors’ reports to extract keywords associated with going-concern issues. Eigenvector analysis is then used to generate category labels, remove irrelevant annotation indicators and identify representative industries. Using a sample of Chinese listed companies from 2017 to 2022, Fisher Score-based dimensionality reduction is applied to optimize input variables and eliminate outliers. A random forest model is subsequently used for prediction, with comparative analyses conducted using backpropagation (BP) neural networks and logistic regression models.
The random forest model achieves a prediction accuracy of 98.71%, demonstrating strong predictive capability. Comparative results further indicate that optimized input features significantly improve the performance of BP neural networks and logistic regression models. The findings suggest that combining social network analysis with machine learning can effectively enhance the accuracy and interpretability of going-concern predictions.
This study contributes to the literature by integrating SNAML into a unified framework for going-concern prediction. It also provides an interpretable classification approach for identifying firms with elevated going-concern risks and demonstrates the practical applicability of machine learning techniques in auditing and risk assessment.
