Deep Learning for Stock Market Prediction: A Systematic Review

Stock market fluctuations is challenging due to nonlinear price dynamics and short-term investor sentiment effects. Deep learning models address challenges by combining historical price and trading/volume features with sentiment signals from financial news, social media, or blended sources. Although studies report strong results, evaluation practices remain inconsistent, ignoring real world viability. This systematic review analyses 22 papers published 2015-2024 that integrate historical and sentiment data for stock market forecasting. For classification, accuracy is predominant, reported in 13 of 22 studies. However, only 6 papers supplement raw accuracy with discriminatory performance measures, limiting confidence in directional reliability. Among regression papers, 4 report root mean square error (RMSE) and 4 report mean absolute error (MAE) as headline metrics. 5 studies report finance-aware metrics, including net profit or Sharpe ratio; 1 paper adjusts for transaction fees. No regression study reports finance-specific metrics such as maximum drawdown or transaction cost evaluation. Findings show sentiment integration benefits the next-day time horizon; 10 studies report a median accuracy gain of +7.4%. Weaker persistence at longer horizons suggests sentiment accuracy decay. Datasets vary from millions of sentiment records to small annotated samples, often without clear documentation, reducing reproducibility. Sentiment enhanced deep learning improves predictive metrics relative to price only baselines; however, minimal cost aware evaluation and real-time deployment analysis prevent substantiation of commercial viability. Rigorous, standardised, finance-aware evaluation frameworks are required to reconcile academic benchmarks with deployable trading systems.