A Systematic Review of Machine Learning Approaches for Medium Horizon Stock Direction Prediction in Emerging Markets

Machine learning (ML) methods are widely applied to stock market prediction, and many studies report that ML models can capture nonlinear patterns and sometimes outperform simple benchmarks. However, most of this work focuses on very short horizons such as intraday or next‑day movements, while medium‑horizon prediction over one to six months remains comparatively underexplored, particularly in emerging markets. Existing studies show that both classical ML models and deep learning architectures can extract useful patterns from prices, volumes, fundamentals, and macro indicators, but reported performance is highly sensitive to the prediction horizon, market type, and evaluation protocol. At the same time, recent reviews and methodological analyses highlight common problems such as unrealistic random train–test splits, data leakage, and lack of reproducibility, which make it difficult to judge how reliable these models truly are. This review provides a concise overview of ML‑based methods for stock direction prediction with an emphasis on medium‑term horizons and emerging markets. It clarifies the prediction problem and horizons, summarizes typical data sources and feature families, outlines the main model types used in the literature, and discusses evaluation practices and pitfalls. The review then identifies open challenges related to non‑stationarity, regime shifts, and real‑world deployability, and outlines directions for more standardized and reproducible research on medium‑horizon stock direction prediction in emerging markets.