<p>Accurate stock price prediction remains challenging due to the market’s nonlinear dynamics and behavioral complexity. Existing models often neglect retail investors’ sentiment and its amplification through popularity. This study proposes XSI-BiLSTM, a hybrid deep learning framework that improves stock price forecasting by incorporating retail investor sentiment. In this framework, we pioneer the use of XLNET to extract investor sentiment from stock market forums and, for the first time, integrate the resulting popularity-weighted sentiment indices with a BiLSTM-highway network for prediction, enabling the model to capture complex behavioral and temporal dependencies. Applied to the Chinese financial market, XSI-BiLSTM surpasses benchmarks by at least 8.15% (RMSE), 6.21% (MAPE), and 5.80% (R2). Validation and ablation experiments collectively confirm the contributions of popularity-adjusted sentiment and the highway mechanism to predictive performance of XSI-BiLSTM, jointly enhancing the reliability of forecasts. The findings empirically prove the predictive value of popularity-amplified sentiment and the potential of combining behavioral indicators with deep temporal learning to improve forecasting.</p>