Published Online:July 2026
Product Name:The IUP Journal of Applied Finance
Product Type:Article
Product Code:IJAF050726
DOI:10.71329/IUPJAF/2026.32.3.93-104
Author Name:Rupali Mangesh Sathe, Neetu Sharma and Shilpa Shinde
Availability:YES
Subject/Domain:Finance
Download Format:PDF
Pages:93-104
Rising volumes of digital payments have made financial fraud one of the most pressing concerns for banks, regulators, and consumers alike. This paper reports a head-to-head comparison of two deep learning systems, each enhanced by a distinct metaheuristic optimizer, built specifically for anomaly detection in financial time series. The first system couples a deep convolutional capsule autoencoder (DCCAE) with the adaptive rain optimization algorithm (AROA), which handles feature selection before training begins. The second pairs a graph convolutional long short-term memory network (GC-LSTM) with the Lévy-flight distributed dung beetle optimizer (LFDBO) for continuous hyperparameter refinement. Both the systems are tested on a widely-used European credit card fraud dataset and on a synthetic sequential transaction dataset, applying the same preprocessing steps and a fivefold cross-validation scheme throughout. The GC-LSTM variant delivered a stronger outcome, with more accuracy, a F1-score, and a AUCROC, while also converging faster. A structured ablation exercise confirmed that neither optimizer is dispensable. The findings offer concrete guidance for practitioners selecting fraud detection infrastructure, with direct implications for financial regulators, compliance teams, and fintech product developers.
The accelerated shift to digital financial services has led to an enormous rise in online transactions. While this has enabled quicker and more inclusive financial access, it has simultaneously created fertile ground for increasingly sophisticated fraudulent activities that threaten both system integrity and consumer trust (Craja et al., 2020).