Artificial Intelligence and Machine Learning in High-Frequency Trading: Evidence, Validation Risks, and a Research-Realism Framework

High-frequency trading (HFT) research increasingly uses machine learning to predict short-horizon market movements and reinforcement learning to make sequential trading decisions. Yet strong predictive metrics do not automatically imply executable profitability. This paper revisits an undergraduate research topic on AI and ML for HFT and rebuilds it as an evidence-based review and methodological critique. It synthesises research on support vector machines, limit-order-book deep learning, LSTM/CNN architectures, reinforcement learning, market making, order-flow imbalance, adverse selection and backtest overfitting. The review distinguishes three layers of evidence: prediction, trading decision and execution. It proposes a six-part research-realism framework covering data provenance, temporal validation, trading frictions, execution realism, market interaction and reproducibility. The evidence indicates that ML can learn useful structure from high-frequency order-book data, but credible trading claims require strict chronological out-of-sample testing, cost and fill modelling, sensitivity analysis and controls against repeated strategy selection. The paper concludes with a practical protocol for future empirical work and a discussion of quantitative-finance capacity in emerging financial systems. Disclosure: This is an independent author-revised research paper derived from themes developed in the author's undergraduate thesis at Lincoln University College, Malaysia. It is not peer reviewed and does not present new empirical backtest results. Lincoln University College is not represented as the publisher or endorser of this 2026 revision.