Financial fraud is rising in the digital economy, causing major losses and weakening trust in banking systems. Conventional detection methods struggle with transaction data due to extreme class imbalance, high volume, and complexity, leading to delayed or inaccurate detection. This study proposes an integrated fraud-detection framework using LSTM-GRU (deep learning) and XGBoost (machine learning), supported by rigorous preprocessing: SMOTE for class balancing, outlier handling, normalization, and missing-data/feature treatment. Evaluation using accuracy, precision, recall, F1-score, confusion matrices, and ROC curves shows LSTM-GRU performs best (Acc 98%, Prec 95.19%, Rec 99%, F1 97.06%), outperforming XGBoost and traditional models (SVM, RF, KNN, GAT).