Synthetic Tabular Data for Fraud Detection: A TSTR Comparison of Rule-Based, TVAE, and CTGAN Generators

Synthetic data offers a privacy-mitigating route to training anomaly detectors when labeled transactions are scarce, delayed, or barred from cross-institutional sharing. This study reports a controlled comparison of three tabular synthesizers, a rule-based simulator, a variational autoencoder (TVAE), and a conditional generative adversarial network (CTGAN), under the Train-on-Synthetic, Test-on-Real (TSTR) protocol. Each synthesizer is paired with three training strategies: standard supervision, weak supervision under simulated label delay, and semi-supervised fine-tuning with a small real sample, the last a controlled relaxation of strict TSTR. Detectors are evaluated on the European cardholder dataset released by the Université Libre de Bruxelles, with synthetic-to-real transferability assessed through a controlled distribution-shift sweep; IEEE-CIS Anomaly Detection and Bank Account Anomaly are used to characterize the cross-corpus statistical gaps that motivate the sweep rather than as separate retraining targets. CTGAN attains the closest statistical fidelity to real transactions, with a mean Kolmogorov--Smirnov distance of 0.092, and the strongest TSTR F1 of 0.68 when combined with semi-supervised fine-tuning, recovering roughly 82% of the real-data upper bound of 0.83. Weak supervision degrades more gracefully than standard supervision as the label-delay window widens from zero to fourteen days. Reliable transfer holds up to roughly a 10% synthetic-to-real statistical gap, beyond which all synthesizers lose F1, and the learned generators relinquish their advantage over the rule-based floor only as the gap approaches 40%. The results indicate that synthetic training is a workable but bounded substitute for real labels, with moderate rather than decisive gains, and that generator fidelity and a modest real anchor jointly govern transferability.