Temporal Graph Networks for Bank Customer Churn Prediction with Dynamic

Peihao Sun

2025

Abstract

Customer churn prediction is a critical issue in banking, with a direct influence on customer retention strategy and financial health. Traditional churn models such as Recency-Frequency-Monetary (RFM) analysis ad static graph-based models fail to capture the dynamic nature of client-product interaction, social influence effects, and policy shocks. To address these limitations, this work proposes a Temporal Graph Network (TGN)-based model that integrates dynamic node embeddings and causal temporal difference mechanisms to model real-time financial transactions. The proposed approach is tested on a 12-month transactional dataset of a European retail bank, with a 12–18% improvement in prediction accuracy over baseline models such as LSTM, GCN, and Node2Vec. By leveraging TGN-computed churn risk scores, the paper applies three levels of tailored retention interventions, i.e., dynamic fee remissions, referral stabilization rewards, and policy-sensitive rate adjustments, which collectively boost customer lifetime value (CLV) by 14% and cross-sell rates by 22%. The findings show the effectiveness of temporal graph-based modeling in financial analytics, presenting an interpretable and scalable churn prediction solution. Subsequent research must explore federated learning techniques to enable privacy-sustaining cross-bank collaboration as a supplement to the impact of temporal graph-based knowledge on financial decision-making.

Paper Citation

in Harvard Style

Sun P. (2025). Temporal Graph Networks for Bank Customer Churn Prediction with Dynamic. In Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE; ISBN 978-989-758-765-8, SciTePress, pages 324-331. DOI: 10.5220/0013688900004670

in Bibtex Style

@conference{icdse25,
author={Peihao Sun},
title={Temporal Graph Networks for Bank Customer Churn Prediction with Dynamic},
booktitle={Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE},
year={2025},
pages={324-331},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013688900004670},
isbn={978-989-758-765-8},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 2nd International Conference on Data Science and Engineering - Volume 1: ICDSE
TI - Temporal Graph Networks for Bank Customer Churn Prediction with Dynamic
SN - 978-989-758-765-8
AU - Sun P.
PY - 2025
SP - 324
EP - 331
DO - 10.5220/0013688900004670
PB - SciTePress