Dynamic Weighted Multimodal Financial Forecasting Models: Fusion Strategies and Market Validation

The complexity and multi-source heterogeneous nature of financial markets pose challenges to traditional prediction models, where a critical challenge lies in effectively integrating multi-modal data and dynamically capturing the evolution of market states. Addressing the limitations of existing financial prediction models that rely on static weighting or singlemodal data, this paper proposes a Dynamic Weighting Multimodal Fusion (DWMF) framework. The model introduces an attention mechanism-based dynamic weighting allocation module to analyze real-time information contributions from multi-modal data, including market volatility, investor sentiment, and macroeconomic events. It adaptively adjusts fusion weights of different modalities in time-series forecasting. Furthermore, adversarial training strategies are incorporated to enhance feature representation robustness, while Gated Recurrent Units (GRUs) are employed to capture cross-modal temporal dependencies. Empirical studies on the A-share and U.S. stock markets validate the framework through multi-scenario tests using the CSI 300 Index, S&P 500 constituent stocks, and social media sentiment data. Results demonstrate that compared to static fusion models and single-modal benchmarks, DWMF reduces prediction errors by an average of $25.4 \%$, improves the risk-adjusted return (Sharpe ratio) by $\mathbf{3 8. 5 \%}$, and exhibits superior robustness during extreme market fluctuations. Additionally, a coupling effect is identified between enhanced pricing efficiency and volatility suppression, with the model achieving $83 \%$ of information pricing within one hour following a policy shock. This study establishes a dynamic decision-making paradigm for multi-modal financial data fusion, while its modular architecture demonstrates scalability for cross-market and cross-asset prediction tasks.