Machine learning-based risk prediction of corporate digital transformation

Abstract

Digital transformation creates long-term opportunities for firms, but it may also generate short-term financial pressure during implementation. Existing studies mainly examine the ex-post effect of digital transformation on firm performance, while relatively few focus on Ex-Ante risk prediction. To address this gap, this study develops a machine learning-based framework to predict next-period financial deterioration risk in the context of corporate digital transformation. A binary risk event is defined as a case in which next-year return on assets falls into the bottom 25% within the same calendar year. Based on firm-level variables observed in the current period, this study compares Logistic Regression, Random Forest, and Gradient Boosting Decision Tree models under firm-grouped cross-validation and out-of-time validation. The results show that next-period financial deterioration risk is meaningfully predictable. When current ROA is included, all three models achieve stable performance, with out-of-time AUC values around 0.82 and PR-AUC values around 0.61. Model performance declines substantially when current ROA is excluded, indicating that historical profitability is a key anchor for short-term risk identification. At the same time, digital transformation and related firm characteristics provide useful incremental information. The nonlinear models perform competitively, but they do not consistently outperform the logistic regression benchmark. This study shifts the analytical focus from ex post performance evaluation to ex ante risk prediction and provides a practical basis for early warning in corporate digital transformation.