Predicting customer transaction behaviour in financial industries remains a complex challenge due to limitations in traditional statistical methods, which often struggle with high-dimensional data and fail to account for dynamic external factors. Moreover, existing AI-based prediction models face issues of data biases, limited interpretability and problems with human-crafted features. This study proposes an enhanced AI-based transaction prediction framework, integrating a Feedforward Neural Network (FFNN) with Bayesian optimisation for hyperparameter tuning. The model is trained on the Santander customer transactions dataset, focusing on binary classification of transaction outcomes. The network architecture includes 202 input neurons, two hidden layers of 10 neurons each using ReLU activation, and an output layer with sigmoid activation for probabilistic classification. To improve generalisation and stability, a dropout rate of 0.2 and the Adam optimiser are employed alongside binary cross-entropy loss. Bayesian optimisation is applied to fine-tune parameters such as learning rate, batch size, and dropout rate. The model achieved an accuracy of 0.93, outperforming benchmark models including decision trees, random forests, logistic regression, and gradient boosting. In addition, time-aware metrics such as the Time-Weighted F1 Score (0.52) and Temporal ROC-AUC (0.75) highlight its ability to adapt over time. The architecture is well documented and designed to enhance transparency, mitigate bias, and support practical deployment in financial transaction prediction systems.