Reliable Stock Prediction: Data, Models, Testing

Abstract

In recent years, deep learning and large language models have entered almost every discussion on stock price prediction. Many reported results look strong on paper, but often rely on clean data, cheap trading, and generous assumptions that rarely hold in real markets. This review looks at studies from 2020–2025 through three practical lenses. First, data and task design: how prices, order books, and news are collected, filtered, labeled, and aligned with the information actually available at decision time. Second, models and multimodal methods: long-horizon forecasters, order-book based models, and text–market fusion schemes, all compared against simple but competitive baselines instead of strawman benchmarks. Third, evaluation and implementation: temporal cross-validation, tests for backtest overfitting, explicit treatment of transaction costs and liquidity, and the engineering and compliance constraints of deployable systems. Taken together, these perspectives argue that credible stock prediction work depends less on one more novel architecture and more on transparent data pipelines, honest testing, and designs that could survive live trading.