A two-stage hybrid framework that integrates internal default prediction models with external credit ratings at the decision-making level is proposed, demonstrating that decision-level integration of internal and external models, and addressing class imbalance, enhances both predictive performance and profitability.
Abstract
With the rapid growth of online lending platforms, credit risk management has become increasingly important. Aligned with Basel Committee recommendations, this study proposes a two-stage hybrid framework that integrates internal default prediction models with external credit ratings at the decision-making level. Unlike prior studies that either rely solely on internal models or treat external ratings as input features, the proposed framework preserves the distinct strengths of both sources. In the first stage, twelve machine learning models are combined with five data balancing techniques and feature selection, yielding 60 distinct configurations evaluated under class imbalance. Performance is assessed using conventional metrics (F1, G-mean, AUC) and a profit-based metric (Profit_Score) that reflects the economic impact of model decisions by quantifying avoided losses and forgone revenues. Logistic regression with random oversampling is selected as the optimal model. The key methodological contribution lies in the second stage, where a dynamic credit rating adjustment mechanism is introduced based on a composite score integrating predicted default probability, external credit rating, and loan amount. Results show that the dynamic approach outperforms both the static strategy (by 14.55%) and the standalone internal model (by 30.4%). The findings demonstrate that decision-level integration of internal and external models, and addressing class imbalance, enhances both predictive performance and profitability.
The discussion shows that ensemble methods can provide strong discrimination across public credit datasets, while the usefulness of a model also depends on whether its outputs can be audited and communicated.
Ze-Kun Li· Advances in Economics, Manag...· 0 citations
This study presents a credit risk estimation framework that integrates firm-level financial indicators, bank account activity, and counterparty information derived from transaction data. While recent research has applied Graph Neural Networks (GNNs) to incorporate inter-firm relationships, accumulating evidence suggests that tree-based models often outperform deep learning methods for structured financial data. We propose a tree-based approach that selects economically significant counterparties based on the sales-to-deposit ratio and incorporates their financial and account characteristics into the prediction model. Using a nine-year, large-scale dataset of Japanese firms, we conduct a rolling-window evaluation and compare the proposed model with standard machine learning methods and a GNN-based approach. The empirical results demonstrate that the proposed model achieves superior predictive accuracy while maintaining interpretability, highlighting the practical value of combining account information with structured counterparty features in credit risk assessment.
Yuto Horikoshi, Rei Yamamoto· International Journal of Fin...· 0 citations
An Explainable Machine Learning (XML) framework for credit risk assessment that combines an ensemble classifier, integrating XGBoost, Random Forest, and LightGBM, with an integrated SHAP-and-LIME explainability layer is proposed and evaluated using a large-scale retail and priority-sector loan dataset drawn from public sector, private sector, regional rural, and small finance bank segments operating in India.
A. Agrawal, Vaibhav C. Gandhi· International journal of com...· 0 citations
A weak and unstable relationship between ESG measures and HGK credit ratings is found, and ESG ratings in this context appear to reflect firm size, sector, and reporting capacity more than a strong standalone signal of credit risk.
Vlatka Bilas, Tomislav Radoš, Lana Frkovic· Communications of Internatio...· 0 citations
This research compares the effectiveness of machine-learning and traditional statistical techniques in predicting annual credit rating downgrades for Thai non-financial firms listed on the Stock Exchange of Thailand during 2018–2023, using a time-ordered train-validation-test framework for predictive model evaluation. The machine-learning models are the artificial neural network (ANN), Random Forest, and XGBoost. The traditional techniques are the linear probability model, logistic regression, and probit regression. Model performance is evaluated using threshold-free PR-AUC and threshold-based metrics because downgrade events occur infrequently. The findings show that logistic regression is more useful for risk ranking, while Random Forest is more useful for downgrade screening. Under threshold-free metrics, logistic regression has the highest PR-AUC of 0.214. Under threshold-based metrics, Random Forest performs best, with a recall of 0.455 and an F1 score of 0.233, while the ANN is second-ranked. Overall, the results show that the preferred model depends on the purpose of use. Logistic regression is more suitable when the objective is to rank firms by downgrade risk, while Random Forest is more suitable when the objective is to screen firms for possible downgrade. Given the limited number of downgrade events in the test sample, the model comparisons should be interpreted as indicative rather than definitive.
Jiroj Buranasiri, Prajya Ngamjan, Nuttawaree Ratchpiboon· The Economics and Finance Le...· 0 citations
It is suggested that superior ranking performance does not necessarily imply superior decision quality and that effective credit risk modeling requires balancing predictive flexibility with probabilistic reliability and governance stability.
