Integrating event information and multi dimensional relationships for improved financial time series forecasting

Area of Science:

• Quantitative Finance
• Machine Learning
• Financial Econometrics

Background:

• Financial time series prediction is complex due to market narratives and inter-asset dependencies.
• Traditional models struggle to differentiate price patterns stemming from diverse underlying causes.
• This limits predictive accuracy in real-world financial applications.

Purpose of the Study:

• To develop an advanced deep learning framework for enhanced financial time series prediction.
• To address the limitations of existing models in capturing market narratives and complex asset relationships.
• To propose the Dual-stream Alpha Factor Fusion Network (DAFF-Net).

Main Methods:

• DAFF-Net integrates event-driven temporal pattern extraction using an event-aware router.
• It fuses time series data with contextual event information from news and announcements.
• A multi-dimensional relationship-aware channel soft clustering module captures cross-asset dependencies.

Main Results:

• DAFF-Net demonstrated significant performance improvements over eight baseline models.
• The framework achieved 7.4%-15.2% better Mean Squared Error (MSE) and 7.0%-21.4% enhanced [Formula: see text] metrics.
• DAFF-Net showed particular strength in long-term financial prediction tasks.

Conclusions:

• Integrating event information and multi-dimensional relationships is effective for financial prediction.
• DAFF-Net offers a novel technical paradigm for quantitative investment and risk management.
• The study validates the approach on Amazon stock data and cross-sector stocks.