Evolution and Challenges of Reinforcement Learning in Intraday Algorithmic Trading: A Systematic Literature Review

Reinforcement learning has been increasingly applied to intraday financial decision-making, but the available evidence remains fragmented across heterogeneous tasks, datasets, and evaluation protocols. This systematic literature review examined peer-reviewed studies indexed in ScienceDirect and Scopus from 2018 to 2025. Across the initial search and the subsequent update, 315 bibliographic occurrences were imported; 51 redundant occurrences were removed, leaving 264 unique records for title and abstract screening. Twenty-one reports were retrieved and assessed in full text. Five were excluded during the historical StArt workflow, and 16 were historically accepted. During manuscript revision, these 16 reports were reassessed against the unchanged eligibility criteria; eight were excluded, and eight studies remained in the revised synthesis. The eligible studies address direct stock trading, stock or equity-index volatility forecasting, optimal execution, stock portfolio management, sentiment-informed trading, and dynamic strategy selection. The studies report promising task-specific results, but the evidence is heterogeneous and predominantly based on historical-data experiments or simulation. Transaction costs, execution constraints, and market impact are modeled inconsistently, while differences in objectives, datasets, metrics, baselines, and validation procedures prevent defensible numerical pooling or general ranking of the methods. The available evidence therefore does not establish general superiority, profitability, robustness, or readiness for live deployment. The review is reported in accordance with PRISMA 2020.