GS-FUSE: Granger-Supervised Gated Fusion for Event-Driven Financial Forecasting
GS-FUSE, a novel multimodal forecasting framework, tackles the difficulty of anticipating market responses to financial events by strategically combining textual event information with past price signals. Created by a team of researchers, this model features a Granger-supervised gated fusion module that activates for event text solely when it adds predictive value beyond existing prices. Additionally, it utilizes multi-granularity alignment to link broad event representations with detailed textual indicators and future market trends. Designed as a plug-and-play adapter for both large language models and time-series models, GS-FUSE seeks to enhance directional forecasting of event-to-price movements for investors and policymakers.
Key facts
- GS-FUSE is a multimodal event-based forecasting framework.
- It uses Granger-supervised causal-aware gated fusion.
- The gated fusion module learns when event text is predictive beyond historical prices.
- It includes a multi-granularity alignment mechanism.
- The framework aligns event representations and textual cues with market trajectories.
- It is a plug-and-play adapter for off-the-shelf LLMs and time-series models.
- The goal is to forecast the impact of financial events on markets.
- The paper is available on arXiv with ID 2605.28520.
Entities
Institutions
- arXiv