Machine Learning for Financial Market Prediction : A Systematic Literature Review

Abstract

Artificial intelligence has recently become a powerful tool for market stock prediction. Financial markets are complex and volatile; each market behaves differently. The machine learning field hasand inputs known extensive study over the last few years. Researchers have experimented with a range of inputs for basic price prediction, employing basic market data from historical data and technical indicators to market indexes and alternative data sources such as sentiment indicators and news. However, as the goal is to generate forecasts of prices and trends, it is crucial to assess models through adequate metrics. Yet the field suffers from methodological inconsistency and a persistent disconnect between predictive accuracy and economic viability. This systematic literature review employs a structured PRISMA protocol to collect 63 studies and to examine forecasting models, evaluation metrics, alternative data integration, and real-world deployment challenges. Our analysis reveals that Long Short-Term Memory networks appear in 53% of reviewed papers and dominate the literature, and hybrid architectures account for 61% of the total. To bring conceptual clarity, we propose a four-tier taxonomy encompassing data-level, model-level, parallel-ensemble, and architectural hybrid models. The literature reveals critical gaps: 97% of papers use technical metrics while only 9% report the use of profitability, and only 4 studies apply backtesting with realistic transaction costs. Furthermore, survivorship bias is acknowledged in only three papers, while geographical coverage is heavily concentrated in the United States and Chinese markets, leaving African and Latin American markets understudied. Notably, memorizing narrow market regimes rather than generalizing exhibits overfitting risk. We conclude that the assumption of accuracy equaling profitability is unsupported by current evidence, as experiments reporting profitability and backtesting performance are largely absent, which undermines claims about the model's robustness. We advocate for standardized benchmarks, rigorous ablation studies, profitability evaluation protocols, and controlled experiments studying the effect of preprocessing and architectural synergy separately.