Artificial Intelligence-Driven Multimodal Sensor Fusion for Complex Market Systems via Federated Transformer-Based...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Financial Data Analysis

Background:

• Modern trading systems generate vast, heterogeneous data from multiple sources.
• Traditional single-source models struggle with data complexity, temporal scales, and privacy constraints.
• Exploiting multisource information is crucial for accurate market prediction.

Purpose of the Study:

• To propose a federated multimodal prediction framework (FMST) for complex market systems.
• To address limitations of traditional models in handling multisource, heterogeneous market data.
• To enhance prediction accuracy while preserving data privacy.

Main Methods:

• Uniformly modeling diverse data sources as multimodal time series.
• Utilizing a multimodal market-sensor representation for unified feature encoding.
• Employing a cross-modal Transformer fusion architecture for dynamic interaction analysis.
• Implementing a federated collaborative learning mechanism for privacy-preserving optimization.

Main Results:

• FMST significantly outperforms traditional statistical and deep learning models.
• Achieved RMSE of 0.1136, MAE of 0.0832, R2 of 0.8517, and 74.56% direction accuracy in main prediction.
• Demonstrated strong cross-region generalization with RMSE of 0.1242, MAE of 0.0908, R2 of 0.8261, and 72.48% direction accuracy.
• Ablation studies confirmed the contributions of core components: multimodal representation, cross-modal fusion, and federated learning.

Conclusions:

• FMST effectively integrates multisource market information for enhanced prediction.
• The framework offers a new AI-driven pathway for economic data analysis using multimodal sensing.
• Federated learning enables collaborative optimization without compromising data privacy.