Explainable-AI powered stock price prediction using time series transformers: A Case Study on BIST100

Transformer models significantly outperformed traditional approaches for stock price prediction on Turkish bank stocks, suggesting that advanced deep learning architectures can capture complex patterns in financial time series data more effectively than conventional methods.

Explainability techniques (SHAP and LIME) successfully identified which features most influenced predictions, enabling investors to understand why the model made specific forecasts rather than treating it as a "black box"—a critical requirement for responsible AI deployment in high-stakes financial decisions.

Technical indicators combined with transformer models improved prediction accuracy, indicating that enriching AI models with domain-specific financial knowledge (rather than relying solely on raw price data) produces more reliable forecasts for real-world investment applications.

The study emphasizes democratizing financial decision-making through interpretable AI, suggesting that making complex forecasting tools transparent and understandable could help individual investors make more informed choices—though the paper does not address risks of overconfidence or the limitations of historical prediction accuracy.

A 10-year dataset (2015-2025) covering multiple market conditions provides reasonable validation, though the focus on a single market index (BIST100) and specific bank stocks limits generalizability to other markets or asset classes, which users should consider before applying these methods elsewhere.