On Explaining the Sentiments in Prediction of Stock Movement: An XAI-Based Analysis

Empirical evidence shows that short-horizon equity returns are not fully random; there is a degree of predictability, especially when traditional financial data is combined with public sentiment measures. Sentiment-based features have been used to improve prediction tasks, but there is limited explanation of their contribution. In addition, the impact of news on investor sentiment, and how that impact decays over time, has rarely been investigated. To bridge this gap, we integrate sentiment scores extracted from financial news headlines using HKUST FinBERT with daily market-based indicators to predict the next-day price direction. We evaluated two aggregation strategies: the most confident news of the day and average sentiments, and introduced a decay mechanism to attenuate the influence of older news. Predictive performance is benchmarked with an ANN, and SHAP provides model-agnostic feature attribution. Incorporating decayed sentiment from the most confident headline increases the test accuracy from 60.24% (technical indicators only) to 65.06%. SHAP highlights decayed neutral and negative sentiments, overnight sentiments and pre-market adjustments, and weekday effects as the most influential short-term predictors, consistent with prior behavioral finance evidence. These findings underscore the value of sentiment-aware, explainable AI models for short-term forecasting and highlight future improvements by using richer data and enhanced sentiment extraction.