Hybrid Optimized BRB–ML Model for Credit Rating Prediction in Open Banking Systems

This paper presents a hybrid decision support system that integrates a Belief Rule Base (BRB) model, a Machine Learning (ML) model, and Particle Swarm Optimization (PSO). The system predicts the credit rating of fintech clients based on data obtained through the Open Banking API. For a real-world dataset, the classification ML model Random Forest demonstrates high predictive accuracy. The BRB model is represented by a complete set of rules covering four attributes with three referential values for each attribute. By analyzing the distribution of the number of activated rules and their activation weights, the representativeness of the data was assessed. A group of rules with low activation weights was identified, corresponding to the referential value Middle for one of the attributes. PSO was applied to optimize BRB parameters, ensuring both interpretability of the results and the capability to model incomplete data. For model validation, data were synthesized using a Rule-Based Data Synthetic approach. The results show that BRB models maintain stable accuracy across real and validation datasets. In contrast, the accuracy of the ML model decreased on validation data by an average of 5% compared to the initial dataset. These findings highlight the advantages of the hybrid approach, which combines the accuracy of ML with the interpretability and robustness of BRB in data-limited environments