Abstract: Abstract: Stock price prediction plays a critical role in financial markets, serving as a foundation for investment decision-making, risk management, asset allocation, and market evaluation. Stock price fluctuations not only directly impact investor returns but also reflect market responses to macroeconomic trends, policy shifts, and firm-level developments. Although deep learning has achieved notable success in time series forecasting, existing models are often limited to single-modality inputs, neglecting cross-industry correlations and modality heterogeneity, which hampers a comprehensive understanding of market dynamics. While some studies have introduced textual modalities, they typically fall short in addressing challenges such as modality asynchrony and multi-granularity alignment. To overcome these limitations, we propose MATCH, a multimodal fusion framework tailored for financial markets. MATCH integrates a mixture-of-experts pretraining strategy to mitigate modality lag and incorporate industry-specific knowledge. It further employs frequency-domain decomposition and hierarchical fusion mechanisms to enable effective cross-modal interaction under multi-granularity settings. Extensive experiments on real-world stock datasets demonstrate that MATCH significantly outperforms existing state-of-the-art methods in predictive performance.

Key words: financial time series, multimodal, mixture of experts, pretrained model, time-frequency analysis

摘要： 在股票市场中，股价预测在股票市场中具有重要意义，是投资决策、风险控制、资产配置及市场评估的核心依据。股票价格的波动不仅直接影响投资者收益，亦反映市场对宏观经济、行业政策及企业经营状况的反应。尽管深度学习方法在时序预测中取得了显著进展，现有模型多基于单一模态，忽视了行业间的联动效应与信息异质性，难以全面刻画金融动态。部分研究尝试引入文本模态，但在处理模态异构所导致的时滞性、多粒度等问题上仍存在不足，模态对齐与融合方式成为制约多模态预测性能的关键因素。为此，本文提出面向股票市场的多模态融合框架MATCH，提出混合专家预训练策略以缓解模态时滞，注入行业特征信息；结合频域分解与层次化融合机制，实现多粒度场景下的时序与文本有效交互。我们在多个真实股票数据集上的实验结果表明，MATCH在预测性能上显著优于现有主流方法。

关键词: 金融时间序列, 多模态, 混合专家模型, 预训练模型, 时频分析