Fast and reliable data are now essential for lenders that must approve credit, block fraud and update risk positions in milliseconds. Although many machine-learning methods promise to clean and safeguard these data streams, the supporting evidence is scattered and usually tested on a few narrow public sets. This systematic review connects data-quality problems, digital-finance tasks and advanced analytics in a single synthesis. A Scopus search on 23 January 2025 retrieved 4.99 million records; PRISMA filtering reduced this to 1 238 papers that covered all three themes, and full-text screening yielded 133 peer-reviewed studies from 2005–2025. Four technical waves dominate: deep-learning detectors such as Isolation Forest variants, variational auto-encoders and diffusion models; lightweight statistical tests that offer low latency and energy use; data-centric tools that track drift, imbalance and noise in real time; and cross-disciplinary work on fairness, poisoning and privacy. Six gaps persist—over-reliance on a few datasets, little reporting of latency or carbon cost, lack of audit-grade explanations, uncertain realism of synthetic anomalies, poor cross-sector transferability and piecemeal ethical-risk coverage. Addressing these gaps will require privacy-safe longitudinal datasets with energy tags, standard speed-and-carbon metrics, rigorous tests for synthetic data, domain-invariant learning and integrated governance. The synthesis provides researchers with a coherent map of progress and guides practitioners in building anomaly-detection pipelines that are fast, energy-aware and trustworthy.