Credit risk prediction model for listed companies based on improved reinforcement learning and Bayesian optimization...

Area of Science:

• Financial Risk Management
• Machine Learning in Finance
• Computational Economics

Background:

• Credit risk prediction is crucial for financial institutions, regulators, and investors due to rapid financial sector growth.
• Traditional credit risk models (z-score, logit, KVM, neural networks) face challenges in feature selection, imbalanced classification, and hyperparameter optimization.
• Existing methods often yield suboptimal results, necessitating advanced approaches.

Purpose of the Study:

• To develop an advanced credit risk prediction model for publicly traded companies.
• To address limitations of traditional models by integrating reinforcement learning and advanced optimization techniques.
• To enhance the accuracy and efficiency of credit risk forecasting.

Main Methods:

• Utilized an off-policy Proximal Policy Optimization (PPO) algorithm, a reinforcement learning (RL) technique, for effective feature selection and imbalanced classification.
• Employed Bayesian Optimization Hyperband (BOHB) for efficient hyperparameter optimization, merging Bayesian optimization with Hyperband.
• Validated the model on diverse datasets: CSMAR, MorningStar, KMV default, GMSC, and UCICCD.

Main Results:

• The proposed model achieved superior performance compared to state-of-the-art methods across all tested datasets.
• Specific F-measures obtained were 90.763% (CSMAR), 86.358% (MorningStar), 87.047% (KMV default), 90.576% (GMSC), and 89.485% (UCICCD).
• Demonstrated enhanced sample efficiency and optimized data utilization through RL-based PPO.

Conclusions:

• The novel credit risk prediction method shows significant promise and efficiency in financial settings.
• The integration of PPO and BOHB represents a major advancement in credit risk assessment systems.
• The findings support the model's effectiveness in enhancing investigative approaches for credit risk.