APPLICATION OF ARTIFICIAL NEURAL NETWORKS FOR PREDICTING CUSTOMER LOYALTY IN COMMERCIAL BANKS

Abstract

Bank customer behaviour has a pronounced temporal nature: the propensity to churn develops
gradually and manifests itself in the dynamics of transactions, balances, and account activity. Static machinelearning
models that operate on aggregated features ignore the sequence of events and lose significant
predictive information. This paper proposes a customer loyalty prediction model based on a Gated Recurrent
Unit (GRU) recurrent neural network capable of capturing temporal dependencies in customer behaviour. The
model was trained on sequences of monthly behavioural profiles of 10,000 retail-bank customers over 12 months
and compared with a static neural network (MLP), a vanilla RNN, an LSTM network, and logistic regression.
The GRU model achieved ROC-AUC = 0.91 and an F1-score of 0.76, outperforming static approaches and
matching LSTM performance with fewer parameters. The results confirm the effectiveness of accounting for the
temporal dynamics of customer behaviour and the feasibility of applying GRU in customer-retention systems.