Lending is an important activity of financial institutions, providing individuals and businesses with access to funds for different economic needs. The continued growth in lending has also raised concerns about non-performing loans in financial institutions, creating a need for effective measures to mitigate loan default. In Nigeria, loan default prediction is constrained by challenges such as limited access to relevant Nigerian financial data, making it difficult to train models that adequately reflect the local financial environment. This study contributes to loan default prediction by developing a deep learning framework using Nigerian financial data. The framework uses a Bidirectional Gated Recurrent Unit Generative Adversarial Network (Bi-GRU GAN) to generate additional default cases and balance the training data. The balanced data are then passed through an LSTM-based Denoising Autoencoder (LSTM-DAE) to learn a compact representation of the loan features, which is subsequently used by a Bi-GRU classifier to predict loan default. The dataset initially contained 6,030 loan records and was reduced to 6,000 after data cleaning. The data had a default rate of 15.95%, with approximately 5.3 non-default cases for every default case. The Bi-GRU GAN generated 2,860 additional default samples, increasing the default class from 670 to 3,530 and producing a balanced training set. Three model variants were evaluated: Baseline, GAN-Augmented, and Full Framework. The Full Framework achieved the best performance, with 96.3% accuracy, 89.3% precision, 87.4% recall, an F1-score of 88.3%, and a ROC-AUC of 0.941. The results show that combining synthetic data generation, learned feature representation, and Bi-GRU classification improved loan default prediction. The study presents an integrated deep learning framework for loan default prediction.