Overcoming Data Limitations in Credit Risk Assessment with FinGPT-Generated Synthetic Data

In credit risk management, accurately assessing credit risk is critical for the stability and risk mitigation of financial institutions. However, research in this area is often hindered by the limitations of small and imbalanced datasets, which impair the effectiveness of credit classification models. This study addresses the challenge of insufficient financial data by utilizing FinGPT, an open-source large language model, to generate synthetic datasets for credit risk assessment. The study compares the performance of three traditional machine learning algorithms, namely Random Forest, Decision Tree, and K-Nearest Neighbors, trained on both the original and FinGPT-generated datasets. The experimental design involves splitting the dataset into training and testing sets, with different portions used to simulate scenarios of limited data availability. FinGPT, which does not require additional fine-tuning, is evaluated against these traditional models using key metrics including Accuracy, Recall, Precision, and F1 Score. The findings demonstrate that FinGPT can enhance the predictive accuracy of these models, particularly in scenarios where data is scarce, highlighting its potential as a powerful tool for financial applications. This study effectively explored a new possibility in overcoming the limitations associated with small or imbalanced datasets based on the value of synthetic data.