Dynamic weight-adjusted random forest boost: a novel ensemble framework for financial distress prediction in highly imbalanced data

Hou, Guodong (2026) Dynamic weight-adjusted random forest boost: a novel ensemble framework for financial distress prediction in highly imbalanced data. PhD thesis, UTAR.

PDF Download (5Mb)

Abstract

Financial distress prediction is essential for identifying early warning signs of bankruptcy and financial instability. This study aims to improve financial distress prediction by addressing class imbalance, enhancing the learning of persistently misclassified distressed samples, and dynamically adapting ensemble weights. To achieve this, a Dynamic Weight-Adjusted Random Forest Boost (DWARFB) framework is proposed. DWARFB integrates multi-stage sampling, history-aware misclassified samples reweighting, and a dynamic performance-driven ensemble strategy. It improves the detection of distressed samples while preserving the original data distribution through gradual class-ratio adjustment without synthetic data generation. As a result, this approach maintains stability and interpretability in financial applications. An accumulated misclassification-based mechanism is introduced to track persistently misclassified samples across iterations. This mechanism emphasizes borderline but informative observations and mitigating noisy minority effects. Combined with a reward-based sliding-iteration strategy, DWARFB stabilizes training and correctly identifies 85% of the distressed samples. This adaptive ensemble component further stratifies iterative models based on real-time F1 performance, dynamically adjusting weights to optimize generalization in highly imbalanced settings. Empirical findings suggest that the DWARFB framework provides more accurate and reliable predictions than conventional boosting methods and Random Forest models. DWARFB achieves superior balance between precision and recall, higher sensitivity to minority-class distress samples, and robust overall discriminative power. The result findings also confirm the importance of cumulative profitability, retained earnings, equity strength, and earnings stability, aligning with established financial distress indicators. Sensitivity analysis shows that model parameters such as tree depth, leaf size, and Alpha–Gamma weighting jointly govern learning capacity and the precision–recall trade-off, while the number of trees primarily affects efficiency. Overall, DWARFB provides a generalizable, interpretable, and adaptive framework for enhancing tree-based classifiers under severe class imbalance. It offers a practical tool for investors and decision-makers to improve early-warning prediction of financial distress in dynamic corporate environments.

Actions (login required)