In the field of financial risk management, credit scoring is recognized as a vital mechanism for predicting the likelihood of loan repayment by applicants. Although traditional machine learning models have been widely used in this area, the integration of advanced deep learning techniques with ensemble learning paradigms has emerged as a transformative step in enhancing prediction accuracy. This study introduces the Hierarchical Unified Optimal Algorithm (HUOA), which leverages the synergy between three advanced processing layers. At the base layer, three ensemble-based classifiers—AdaBoost, Bagging, and Long Short-Term Memory (LSTM) networks—are employed in parallel to extract primary-level features. The output of this layer is fed into a meta-learning layer, where an adaptive Random Forest meta-classifier performs nonlinear combination of predictions to generate the final credit risk score. In this research, an integrated deep learning framework for bank customer credit scoring is proposed, designed based on ensemble learning and unified optimization of parameters and feature selection using a Genetic Algorithm (GA). Empirical evaluation on data from corporate customers of an Iranian bank, along with the UCI datasets from Australia and Germany, using various metrics—particularly misclassification (MC)—demonstrates that HUOA has achieved significant improvements compared to existing hybrid methods during the 2023-2025 period. This hierarchical architecture not only preserves model interpretability through feature importance analysis in the meta-layer but also provides a robust framework for decision-making in dynamic banking environments by reducing prediction variance in imbalanced class scenarios. Findings indicate that the intelligent integration of LSTM with dynamic sampling strategies in the ensemble layer, combined with adaptive feature selection mechanisms in the meta-layer, can be regarded as a new paradigm in fourth-generation credit scoring systems.