A Closed-Loop Financial Risk Prediction and Dynamic Control Framework Using Optimized Random Forests

Abstract

To address the frequent financial risk (FR) issues exposed by enterprises in the complex economic environment, this paper proposes an optimized model based on random forest (RF), which is used for the early warning and dynamic control of enterprise financial risk (EFR). In existing studies, traditional statistical methods rely excessively on linear assumptions and thus struggle to capture the nonlinear relationships between financial indicators. To tackle these problems, this paper establishes a closed-loop risk management framework of "prediction - classification - control - feedback". Systematic optimization of the RF is further performed, including hyperparameter tuning (e.g., number of trees, maximum depth, minimum samples for splitting, etc.), class weight setting to alleviate the imbalanced sample problem, and indicator selection based on feature importance combined with integration strategy adjustment. This thereby improves the generalization ability and robustness of the model. The optimized model works exceptionally well in terms of prediction accuracy and robustness according to experimental results. For example, on the University of California, Irvine Bankruptcy Prediction (UCI-BP) dataset, its accuracy rate is 0.91, precision rate is 0.88, specificity is 0.90, and Brier score is only 0.11, which is significantly better than the comparison models. Among the robustness indicators, the logarithmic loss is 0.27, Matthews Correlation Coefficient (MCC) is 0.78, Kappa coefficient is 0.75, and stability index is 0.89, all demonstrating stronger robustness. In terms of dynamic control effect, based on the risk probabilities output by the model, this paper classifies enterprises into different risk levels and constructs a strategy mapping rule base of "risk level-control measures". By simulating the implementation of intervention methods such as cash flow optimization and liability structure adjustment, the changes in risk distribution and the transfer rate of high-risk enterprises are evaluated in combination with feedback data. The risk classification accuracy rates of the optimized model on three datasets (UCI-BP, Kaggle Company Bankruptcy Prediction (Kaggle-BP) and Kaggle Corporate Credit Rating (Kaggle-CCR)) are 0.91, 0.89 and 0.88 respectively. Moreover, the transfer rates of high-risk enterprises reach 0.34, 0.32 and 0.30 respectively. According to these findings, the model can successfully assist in the execution of control techniques and precisely identify possible hazards, enabling businesses to take preventative action before crises arise. As a result, this paper offers a fresh approach and useful resource for EFR early warning and control research. It enhances the theoretical framework of risk prediction models and raises the usefulness of risk management in real-world applications.