每日股價的短期預測準確度，可以透過採用半強式市場效率（Semi-strong form efficiency）大幅提升，此方法結合了數值市場數據與公開新聞資訊。我們提出兩種統一的多模態預測框架，將最先進的時間序列模型與預訓練大型語言模型（LLM）嵌入結合起來：方法一將每一種模態分別處理，使用先進的時間序列模型與基於LSTM的文本投影對LLM嵌入進行處理，然後通過交叉注意力機制與加性晚期融合進行結合；方法二則在網絡初期即將所有特徵串接融合。我們在TSMC每日股價資料上進行實驗，配合來自NYSE市場的英文Bloomberg新聞與TWSE市場的繁體中文ETToday新聞。結果顯示，方法一在一至四步的預測中，始終優於單模態（Random Walk, NLinear、iTransformer）與多模態（MM TSFlib）基準模型，展現出最低的預測誤差，而方法二表現則相對較差。消融實驗證實交叉注意力機制能進一步提升預測效能。與通用語言模型相比，經過領域調適的文本嵌入可進一步降低預測誤差，而使用先進大型語言模型進行新聞摘要亦帶來些微改進。本研究結果顯示，透過整合數值與公開新聞文本資料，並採用加性晚期融合、交叉注意力模組、領域專屬文本嵌入與高品質新聞摘要，能實現一種簡潔卻強大的方式，有效提升短期股價預測的準確性與穩定性，展現出半強式效率模型的實際應用潛力。;Accurate short-term forecasting of daily stock prices can be significantly enhanced by adopting the Semi-strong form efficiency, which integrates numerical market data with public news. We propose two unified multimodal frameworks that combine state-of-the-art time series models with pre-trained large language model embeddings: The proposed method 1 processes each modality independently with state-of-the-art time series models, LLM embeddings with LSTM-based text projection, followed by a cross-attention mechanism and additive late fusion, while the proposed method 2 concatenates features early in the network. Experiments on TSMC daily stock data with English news from Bloomberg on the NYSE market and Traditional Chinese news from ETToday on the TWSE market show that the proposed method 1 consistently achieves the lowest one-to-four-step forecast errors against unimodal (Random Walk, NLinear, iTransformer) and multimodal (MM TSFlib) baselines, whereas proposed method 2 underperforms. Ablation studies confirm that cross-attention yields additional gains. Domain-adapted text embeddings further reduce forecasting error compared to general-purpose models, and the news text summarization with advanced large language models delivers marginal improvements. These findings demonstrate that semi-strong from efficiency method by integrating numerical and public news text data using late additive fusion, cross-attention modules, domain-specific text embedding, and high-quality news text summarization constitutes a simple yet powerful approach for improving the accuracy and stability of short-term stock price forecasts.
