GS-FUSE: Granger-Supervised Gated Fusion and Multi-Granularity Alignment for Event-Driven Financial Forecasting
Researchers introduce GS-Fuse, a machine learning framework that improves financial forecasting by intelligently combining event-driven text with price data. The system uses causal analysis to determine when news actually predicts market movements, addressing a key limitation in existing multimodal AI models that treat all data sources equally.
GS-Fuse represents a meaningful advancement in applying artificial intelligence to financial prediction, tackling a fundamental problem in quantitative finance: determining which external events genuinely move markets versus noise. Traditional multimodal models treat text and price data symmetrically, failing to capture the directional relationship where events should primarily inform price predictions, not vice versa. This research bridges that gap by incorporating Granger causality—a statistical method for establishing temporal causal relationships—directly into a gated fusion mechanism. The framework only activates event-driven signals when they provide predictive value beyond what historical prices already reveal, mimicking how experienced traders selectively weight news based on market context.
The broader context reflects growing recognition that large language models and time-series foundation models, while individually powerful, require architectural innovations to work together effectively. Rather than building proprietary end-to-end systems, GS-Fuse functions as a plug-and-play adapter, reducing friction for institutional adoption across different trading systems and asset classes. The multi-granularity alignment mechanism—simultaneously processing event-level semantics and fine-grained textual details—suggests the authors recognized that critical market signals exist at multiple levels of abstraction.
For the financial AI sector, this work demonstrates that sophisticated forecasting increasingly depends on principled fusion strategies rather than brute-force model scaling. Institutions deploying event-driven trading systems could benefit from more reliable signal filtering, though real-world implementation requires handling market microstructure, regime changes, and data quality issues beyond the paper's scope. The research validates that causal reasoning embedded into deep learning architectures offers practical improvements in a domain where accuracy directly translates to competitive advantage.
- →GS-Fuse uses Granger causality to determine when event text genuinely predicts prices, avoiding false signals from noise
- →The framework works as a flexible adapter compatible with existing LLMs and time-series models, enabling rapid institutional deployment
- →Multi-granularity alignment mechanism processes both high-level event semantics and fine-grained textual details simultaneously
- →Experimental results show consistent outperformance over baseline models across multiple assets and forecasting horizons
- →The research advances event-driven financial AI by embedding causal reasoning directly into fusion architecture rather than treating modalities symmetrically