The growing complexity and scale of financial and behavioral datasets have posed significant challenges to traditional credit scoring methods, as conventional models often fail to capture temporal dependencies and nonlinear feature interactions. To address this, a Hybrid Long Short- Term Memory (LSTM) network was designed, incorporating temporal features for accurate credit scoring prediction. The model integrates a hybrid loss function combining binary cross entropy for classification tasks and optimization techniques such as Min-Max normalization, Synthetic Minority Oversampling Technique (SMOTE) for imbalanced data handling, Recursive Feature Elimination (RFE) for feature selection, and Principal Component Analysis (PCA) for dimensionality reduction. Performance evaluation was conducted on benchmark datasets including the Credit Risk dataset, which provides both structured and unstructured financial data. Comparative analysis against baseline models such as Random Forest and XGBoost demonstrated the effectiveness of the approach. Furthermore, a self- attention mechanism was incorporated into the LSTM framework to enhance contextual learning by emphasizing critical input features, leading to improved predictive accuracy. Experimental results indicate that the Hybrid LSTM with self-attention achieved superior performance with 89.87% accuracy, outperforming existing machine learning and deep learning techniques in credit score prediction.