Financial anomaly detection is a rare-event problem operating under three simultaneous constraints that few other machine learning applications face at once: extreme class imbalance, adversarial and non-stationary generative processes, and hard latency budgets imposed by payment authorisation windows. The canonical public benchmark—the European cardholder dataset released by the Université Libre de Bruxelles—contains 492 frauds among 284,807 transactions, a positive rate of 0.172%, and this ratio is, if anything, optimistic relative to production systems. Supervised discriminative learning under such priors is fragile: labels are scarce, delayed, and themselves noisy, since a chargeback may arrive months after the transaction it disputes, and fraudsters adapt precisely to the decision boundaries that supervised models expose.

The autoencoder responds to this constellation of constraints by inverting the framing. Rather than learning a boundary between classes, it learns a compressed reconstruction map fitted to legitimate behaviour and treats reconstruction difficulty as a proxy for atypicality; the formal construction, and the sense in which the map is not a generative model, are set out once in Section 3.1 and are not restated elsewhere in this review. Three consequences of that inversion explain the trajectory documented below. Training requires only the majority class, which sidesteps the labelling bottleneck. The anomaly score is continuous in reconstruction difficulty, so a single model supports operating points from conservative alerting to aggressive blocking. And the learned latent code is a reusable representation that downstream classifiers, clustering routines, and ensemble learners can consume—precisely the hybridisation pattern that dominates the corpus described here. The correspondence between reconstruction error and the probability of fraud is, however, an empirical regularity rather than a guarantee, for reasons developed in Section 3.1.

The foundational demonstrations in the financial setting are now nearly a decade old. Pumsirirat and Yan ^[1] trained an autoencoder alongside a restricted Boltzmann machine on credit card and German credit data, establishing the reconstruction error protocol that virtually every subsequent study reuses; the paper remains the single most cited document in the corpus. In the same year, Zamini and Montazer ^[2] combined an autoencoder with clustering rather than classification, and Jeragh and Alsulaimi ^[3] fed the bottleneck representation to a one-class support vector machine, while Chen et al. ^[4] paired a sparse autoencoder with a generative adversarial network to synthesise minority examples. These four 2018 papers already contain, in embryonic form, the four hybridisation strategies—dimensionality reduction, clustering, one-class classification, and generative augmentation—that the field has been elaborating ever since.

Two developments separate the present state of the field from those first demonstrations. The first is volume. The indexed literature grew from four documents in 2018 to thirty in 2025, an estimated 39.4% per year across complete years, with a structural break separating a slow 2018–2022 regime from a rapid 2023–2025 one; roughly two-thirds of everything published on the subject appeared in the final two complete years. The second is diversification. What began as a single reconstruction error protocol applied to card transactions now spans eleven architectural families and reaches accounting and audit records, insurance claims, payment system telemetry, anti-money laundering screening, and on-chain value transfers.

This review describes that body of knowledge as it currently stands. It sets out the formal construction and the design choices it entails (Section 3, Section 4 and Section 5), the recurrent hybridisation patterns and application domains in which it is used (Section 6 and Section 7), the quantitative structure of the literature that supports it (Section 8), and the evaluation and deployment problems that remain open (Section 9, Section 10 and Section 11). The description is grounded in a documented corpus of 85 Scopus-indexed records published between 2018 and 2026, whose construction and analysis are set out in Section 2 and Appendix A. Two different kinds of claim are made from that material, and they carry different evidential weight. Section 3, Section 4 and Section 5 describe the construction, its architectural variants and the scoring, and thresholding choices they entail; that account rests on the primary methodological literature and on the cited studies individually and does not depend on the corpus being a complete enumeration of the field. Section 6, Section 7, Section 8 and Section 9 describe what this particular body of records does—which designs recur, which data are used, and which components of an evaluation are reported—and every statement there is a statement about the 85 records, with its denominator attached. Broader surveys of machine learning for financial fraud already exist and span a wider methodological range: Thivaios et al. ^[5] screen 101 studies drawn from five databases and treat architectures, data modalities, and deployment challenges together, and Chen et al. ^[6] track year-over-year developments in deep learning for fraud detection. This review is deliberately narrower—one primitive, its variants, and the practice that has grown around it—and its emphasis falls on how that practice is evaluated and reported. Consistent with the purpose of a review, the account reports established results, recurrent design practice, and documented reporting habits rather than new experimental claims; where the underlying evidence is weak, or the diagnostics are negative, this is stated explicitly rather than smoothed over.