🤖 AI

GS-FUSE: Granger-Supervised Gated Fusion and Multi-Granularity Alignment for Event-Driven Financial Forecasting

This paper proposes GS-FUSE, a multimodal framework that enhances financial event forecasting by integrating a Granger-supervised gated fusion module to selectively incorporate predictive text signals and a multi-granularity alignment mechanism to better align textual cues with market trajectories, consistently outperforming existing state-of-the-art models.

Original authors: Yang Zhang, En Chun, Ziyun Mao, Yulu Wu, Jun Wang

Published 2026-05-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yang Zhang, En Chun, Ziyun Mao, Yulu Wu, Jun Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather. You have two sources of information:

The Barometer (Price Data): A machine that records wind speed, temperature, and pressure every second. It's great at showing you the current trends.
The News Anchor (Event Text): A person shouting out sudden updates like "A hurricane is forming!" or "The government just changed the tax laws!"

The Problem:
Most current computer models try to blend these two sources together like a smoothie. They mix the barometer data and the news anchor's voice equally, all the time.

The Flaw: Sometimes the news anchor is just repeating old information or shouting nonsense. If the computer mixes that noise into the weather prediction, it gets confused. Other times, the news anchor says something huge (like a surprise rate hike), and the computer needs to listen very closely, but it treats it the same as a boring update.

The Solution: GS-Fuse
The authors of this paper built a new system called GS-Fuse. Think of it as a smart, cautious Traffic Controller standing between the Barometer and the News Anchor.

Here is how it works, using simple analogies:

1. The "Granger-Supervised" Gate (The Smart Filter)

The core innovation is a special "gate" that decides when to let the News Anchor speak.

How it works: Before the gate opens, the system asks a simple question: "If I ignore the news and only look at the barometer, how bad would my prediction be? If I add the news, does my prediction get significantly better?"
The Analogy: Imagine you are driving. You have a GPS (the price data) that usually works perfectly. Suddenly, a radio announcer (the event text) says, "There is a massive traffic jam ahead!"
- If the GPS already shows a jam, the radio is just repeating what you know. The Gate stays closed. You don't need to change your plan.
- If the GPS shows a clear road, but the radio says "Bridge is out!", the Gate opens wide. You immediately trust the radio because it gives you new, life-saving information that the GPS didn't have.
The Result: The system only listens to the news when the news actually adds something new and useful. It ignores the "noise."

2. Multi-Granularity Alignment (The Highlighter)

Once the gate opens and lets the text in, the system needs to know which part of the text matters.

The Problem: Financial news reports are long. They might have 500 words, but only one sentence (e.g., "Inflation is rising faster than expected") actually moves the stock market. The rest is just boring legal jargon.
The Solution: The system acts like a Highlighter Pen. Instead of treating the whole news article as one big blob, it zooms in. It matches specific words in the text (like "rate hike") with specific moments in the market data (like a sudden price drop).
The Analogy: It's like a teacher grading a student's essay. Instead of just giving the whole essay a grade, the teacher circles the one brilliant sentence that proves the point and ignores the fluff. This helps the computer understand exactly why the market moved.

3. The "Plug-and-Play" Adapter

The authors built this system so it can work with any existing "brain" (Large Language Models for text and Time-Series models for numbers).

The Analogy: Think of GS-Fuse as a universal power adapter. You can plug it into a Samsung phone, an iPhone, or a generic charger. It doesn't matter which specific "brain" you use; the adapter ensures they all work together smoothly to solve the problem.

What Did They Find?

The team tested this system on real financial data, including the S&P 500 (the stock market) and Treasury bonds (government loans).

The Result: GS-Fuse consistently beat all other models. It made fewer mistakes in predicting how prices would move after big news events.
Why it won: It didn't just guess; it learned to be skeptical. It knew when to ignore the news and when to pay close attention, and it knew exactly which words in the news were the most important.

In Summary:
GS-Fuse is a financial forecasting tool that doesn't just "read" the news and "look" at the charts. It acts like a wise investor who knows when to ignore the noise and when to pay attention to the signal, ensuring that every piece of information used to make a prediction actually helps.

Technical Summary: GS-Fuse

Problem Statement

Accurately forecasting the impact of salient financial events (e.g., macroeconomic releases, policy decisions) on asset prices is critical for investors and policymakers. While recent advances in Large Language Models (LLMs) and Time-Series Foundation Models have improved unimodal forecasting, existing multimodal approaches for event-driven financial forecasting face three primary limitations:

Event-Invariant Fusion: Current methods typically fuse text and price data symmetrically using shared layers, assuming a stable balance between modalities. This fails to adapt to the heterogeneous nature of financial markets, where some events are driven by sentiment while others follow pre-existing technical patterns.
Lack of Causal-Aware Fusion: Existing models lack explicit mechanisms to determine when event text provides incremental predictive value beyond historical prices. They often symmetrically fuse modalities, risking the overuse of noisy or uninformative text.
Coarse-Grained Alignment: Standard alignment mechanisms operate at the instance level (global event embedding vs. global time-series embedding). This ignores the fact that only specific passages in financial releases (e.g., forward guidance in FOMC statements) drive market reactions, leading to diluted signals from irrelevant content.

Methodology

The authors propose GS-Fuse, a multimodal framework designed to address these limitations through a Granger-supervised, causal-aware gated fusion mechanism and multi-granularity alignment. The architecture consists of four main components:

1. Encoders

Text Encoder: Utilizes a pre-trained LLM (e.g., LLaMA, Phi-3) to encode event scripts ( $E_i$ ) into token-level representations, which are projected into a shared embedding space.
Time-Series Encoder: Utilizes a pre-trained Time-Series Foundation Model (e.g., MOMENT, Kronos) to encode pre-event market segments ( $X_i$ ) into step-wise representations.
Freezing Strategy: The backbone LLM and TS foundation models remain frozen; only projection heads and downstream modules are trained.

2. Multi-Granularity Cross-Modal Alignment

To bridge the modality gap, GS-Fuse enforces alignment at two levels:

Instance-Level Global Alignment: Uses bidirectional cross-attention to interleave token and step representations, followed by an InfoNCE contrastive loss to align global event and time-series embeddings.
Token/Step Fine-Grained Alignment: Aligns individual text tokens with specific time-series steps. A salience scorer identifies top- $K$ "anchor" tokens, and a contrastive loss encourages these salient tokens to align with their corresponding temporal patterns while repelling unrelated steps from other instances.

3. Granger-Supervised Causal Gating

This is the core innovation. The model employs a gating mechanism that decides whether to incorporate event text based on its incremental predictive utility:

Granger Utility Signal: During training, the model compares a "restricted" predictor (using only time-series data, $\hat{Y}_{ts}$ ) against an "unrestricted" predictor (using both text and time-series, $\hat{Y}_{full}$ ).
Supervision: The loss differential ( $\Delta_i = \ell(Y_i, \hat{Y}_{ts}) - \ell(Y_i, \hat{Y}_{full})$ ) serves as a proxy for Granger causality. If incorporating text reduces the forecasting error, the text is deemed useful.
Gate Training: A gating network predicts a scalar openness score ( $\alpha_i$ ) for each instance. This score is supervised via a distillation loss to match a soft "responsibility" target ( $r_i$ ) derived from the Granger utility. The gate opens ( $\alpha_i \approx 1$ ) only when text provides measurable value and closes otherwise.

4. Decoder

A lightweight Transformer decoder takes the fused representation (weighted by the gate) and generates the future market trajectory over the forecasting horizon $H$ .

Training Scheme

The model is trained in three stages for stability:

TS-only Pre-training: Training the time-series branch to establish a strong baseline.
Text-only Warm-up: Training the text branch to ensure event representations are forecasting-relevant.
Multimodal Fine-tuning: Jointly training the fusion and alignment modules with a combined objective of forecasting loss, alignment loss, and gate supervision.

Key Contributions

Granger-Supervised Gated Fusion: A novel module that learns to selectively trust event text only when it provides incremental predictive value beyond historical prices, effectively modeling the causal direction from event to price.
Multi-Granularity Alignment: A joint alignment scheme that operates at both the instance level and the fine-grained token/step level, allowing the model to localize specific textual cues driving market reactions.
Plug-and-Play Framework: A model-agnostic adapter that can wrap arbitrary LLMs and time-series foundation models, making it flexible across diverse backbones and market settings.

Experimental Results

The authors evaluated GS-Fuse on a curated dataset integrating macroeconomic releases (CAMEF) and financial news (FNSPID) against five assets (S&P500, NASDAQ, INDU, USGG1M, USGG5YR) across three forecasting horizons ( $H \in \{35, 70, 140}$ 5-minute steps).

Performance: GS-Fuse consistently outperformed state-of-the-art unimodal baselines (ARIMA, DLinear, iTransformer, FEDformer) and multimodal baselines (GPT4MTS, TimeCMA, TEST). It achieved the lowest MSE and MAE across all assets and horizons.
Ablation Studies: Removing the Granger gate or the multi-granularity alignment resulted in significant performance drops, confirming that both components are essential and complementary.
Backbone Generality: The framework demonstrated robustness across different LLM (LLaMA vs. Phi) and TS foundation (MOMENT vs. Kronos) combinations, though specific pairings yielded marginal variations in performance.
Causal Analysis: The average gate openness was roughly 51% for equity indices but higher (60.68%) for Treasury yields, indicating the model dynamically adjusts its reliance on text based on the asset class and the specific utility of the event narrative.

Significance and Claims

The paper claims that GS-Fuse addresses the fundamental flaw in existing multimodal forecasting: the assumption that text and price data should always be fused symmetrically. By introducing a Granger-supervised gate, the model explicitly learns the conditional dependence of prices on event text, avoiding the noise of uninformative narratives. The multi-granularity alignment further ensures that the model focuses on the specific parts of an event script that drive market movements.

The authors position GS-Fuse not just as a performance improvement, but as a methodological shift toward causal-aware and adaptive multimodal fusion in finance. The framework is presented as a flexible, plug-and-play solution that can be instantiated with various foundation models, offering a robust approach for event-driven financial forecasting where the predictive value of information is heterogeneous and non-stationary.