REF-AI: Responsible, Explainable and Fair AI for Stock-Market Prediction-A Systematic Literature Review with an Illustrative Multi-Market Case Study

Saeed, Muhammad Atif, Jamil, Akhtar
Asia-Pacific Financial Markets · Springer Nature · 2026-11-03 · DOI: 10.1007/s10690-026-09652-9

Аннотация

Stock-market prediction is a difficult, multidisciplinary problem because financial time series are nonlinear, volatile and non-stationary. This study presents a Systematic Literature Review (SLR) of artificial-intelligence techniques for stock-price prediction, covering statistical, machine-learning, deep-learning, continual-learning, reinforcement-learning, multimodal and ensemble methods, through the lens of Responsible, Explainable and Fair AI (REF-AI). Following PRISMA 2020, 140 studies published between 2018 and July 2026 were included (139 peer-reviewed publications and one arXiv preprint) and coded by method family, task, data modality, market and REF-AI dimension. The review finds rapid growth in model complexity but concentration on a few markets, heavy reliance on price data and price-level accuracy metrics, and limited reporting of explainability (20 studies), uncertainty (11) and fairness (6). A case study then evaluates 23 models on eight stocks, two from each of the United States, Pakistan, India and Malaysia, using next-day return forecasts, an expanding-window walk-forward design, out-of-sample $$R^2$$ , directional-accuracy, Diebold–Mariano, Clark–West and Model Confidence Set tests, and a transaction-cost-adjusted trading exercise. Out-of-sample return predictability is weak for all model families: median out-of-sample $$R^2$$ values are slightly negative, directional accuracy is close to 50%, and the random walk remains in the Model Confidence Set for every stock, while all models, including the random walk, achieve price-level $$R^2$$ above 0.97. The main exception is a short-horizon reversal in the two Malaysian stocks, whose trading value is sensitive to costs. Conformal intervals are close to nominal coverage overall but under-cover in bear regimes, and SHAP, integrated-gradients and attention analyses reveal stock-specific attributions that are informative only where a predictive signal exists. Because each market is represented by two stocks, the case study is illustrative and not a general test of developed versus emerging markets. The findings indicate that no model family is universally superior, that predictive and economic claims require return-based, benchmark-relative and cost-adjusted evidence, and that explainability, uncertainty quantification and regime-disaggregated evaluation are essential for responsible deployment.

Читать у издателя →

Похожие статьи