In a market environment characterized by rapid multi-source information diffusion, relying solely on historical trading data may be insufficient to capture certain dimensions of stock price fluctuations. This is especially true for textile and apparel firms, whose stock prices are highly sensitive to public opinion and investor sentiment due to their strong consumer linkage and brand sensitivity. Against this backdrop, this study develops a multimodal prediction framework that integrates market trading data with textual sentiment features, using a Long Short-Term Memory (LSTM) network to forecast stock movement direction. Price variables and technical indicators are extracted from trading data, while sentiment features are derived from financial news, corporate announcements, and investor comments. These heterogeneous data sources are temporally aligned and fused at the trading-day level. On this basis, traditional machine learning models, unimodal LSTM models, and a multimodal LSTM model are constructed. The predictive value of textual sentiment is systematically examined through comparative experiments, ablation analysis, and robustness tests. The findings show that historical trading data provide an important basis for prediction, while textual sentiment features offer additional explanatory information and enhance the model's ability to identify short-term price direction. Compared with unimodal and conventional benchmark models, the multimodal LSTM model combining trading features with sentiment features delivers superior overall predictive performance. This study provides new empirical evidence for stock price prediction in an industry-specific setting and offers useful implications for investment decision-making, public opinion monitoring, and corporate market risk management.

textile and apparel firms, stock price fluctuation prediction, multimodal sentiment analysis, LSTM