Innovative Credit Risk Assessment System Using Artificial Intelligence

This research presents an explainable deep learning framework for credit risk assessment that integrates an unsupervised autoencoder with a GRU–LSTM hybrid model for sequential data classification. Traditional credit scoring systems face several challenges, including difficulty in capturing complex borrower behavioral patterns, methodological limitations, and a lack of explainability, which reduces their suitability for decision-making in dynamic financial environments. The proposed framework employs an autoencoder to preprocess data by reducing dimensionality and noise, while the GRU–LSTM architecture captures both short- and long-term dependencies in borrower behavior. The autoencoder performs dimensionality reduction by compressing high-dimensional input features into a smaller latent representation through an encoder network trained to reconstruct the original data with minimal information loss. This process removes redundant and noisy attributes while preserving the most informative patterns required for accurate credit risk classification. To support interpretable credit risk evaluation, SHapley Additive exPlanations (SHAP) is used to provide both local and global feature importance explanations. By quantifying the contribution of each input feature to individual predictions, the explainability component supports transparent credit decisions and enables financial institutions to justify automated outcomes within regulatory and auditing workflows. The framework was implemented using Python and evaluated on the German Credit dataset after preprocessing with one-hot encoding and Min–Max normalization. Experimental results demonstrate strong performance, achieving an accuracy of 99.12%, precision of 98.82%, recall of 98.78%, and an F1-score of 98.851. These findings indicate that the framework provides an accurate, explainable, and real-time approach to credit risk assessment in institutional financial settings.