Deep Learning Algorithms for Customer Behaviour Prediction and Recommendation in Banking: A Technical and Algorithmic Comparative Study

Abstract

This paper provides a focused technical and algorithmic treatment of the deep learning architectures underlying modern customer behaviour prediction and recommendation systems in banking, complementing the applied systems-and-governance perspective developed in a companion paper. We formalize the sequential customer-behaviour prediction problem and present the mathematical foundations of five neural architectures  recurrent networks (LSTM, GRU), temporal convolutional networks, the Transformer self-attention mechanism, and autoencoder-based representation learning — together with a proposed hybrid fusion architecture that combines sequential and representation-learning signals for joint behaviour prediction and recommendation ranking. We further formalize the recommendation-ranking objective through matrix factorization and Bayesian personalized ranking, describe the training algorithms (backpropagation through time, the Adam optimizer, and regularization strategies) used to fit these models, and provide a formal computational complexity analysis comparing time and space costs across architectures as a function of sequence length and model width. Comparative experimental analysis examines the accuracy-parameter efficiency frontier, convergence speed, and the interpretability of learned self-attention weights, and a decision framework is offered to guide architecture selection under differing latency, interpretability, and data-volume constraints. The analysis indicates that the proposed hybrid architecture achieves the strongest accuracy-per-parameter efficiency among the compared approaches, at a quadratic-in-sequence-length computational cost inherited from its Transformer component, a tradeoff that institutions should weigh explicitly against latency requirements and customer-history length.