CAMEF: Causal-Augmented Multi-Modality Event-Driven Financial Forecasting by Integrating Time Series Patterns and Salient Macroeconomic Announcements

Yang Zhang, Wenhan Yang, Jun Wang, Qiang Ma, Jie Xiong

TL;DR

This paper introduces CAMEF, a novel framework for predicting financial market movements driven by macroeconomic announcements. By integrating textual analysis of announcements with historical time-series data and a unique causal learning mechanism augmented by LLMs, CAMEF significantly improves forecasting accuracy, outperforming existing methods.

What Problem Does This Paper Address?

Forecasting the impact of macroeconomic events on financial markets is crucial for investors and policymakers. These events, such as interest rate decisions or employment reports, can trigger significant market shifts by influencing economic expectations. While existing methods often focus on either textual analysis of announcements or time-series modeling of market data, they struggle to capture the complex, multi-modal nature of financial markets and, critically, the causal relationships between events and market reactions. Furthermore, many studies rely on less granular daily data, limiting their applicability in high-frequency trading scenarios. This paper aims to bridge this gap by developing a more comprehensive, causal, and multi-modal approach using high-frequency data.

Methods

CAMEF is a multi-modality framework designed to tackle the limitations of previous approaches. Its core components include:

  • Multi-Modality Integration: CAMEF combines textual information from macroeconomic announcements (processed by RoBERTa) with high-frequency financial time-series data (processed by MOMENT). These modalities are fused using a specifically designed feature fusion layer.
  • Causal Learning: A key innovation is the incorporation of causal learning. This involves generating counterfactual event scripts using Large Language Models (LLMs) to create hypothetical scenarios. These counterfactuals, along with the original event and time-series data, are used in a causal contrastive learning objective (triplet loss) to ensure the model learns genuine causal relationships rather than spurious correlations.
  • LLM-based Counterfactual Augmentation: The framework leverages LLMs (specifically LLaMA-3) to generate diverse counterfactual event scripts. This process involves summarizing original texts, analyzing their sentiment, and then modifying key facts and sentiment-related phrases to create altered versions that reflect different target sentiments, while preserving structural consistency.
  • Time-Series Forecasting Decoder: A GPT-2 based decoder is used to generate future time-series predictions based on the fused multi-modal representation. A post-regressor further refines these predictions.

Datasets

The paper introduces a novel, open-source synthetic dataset specifically curated for event-driven financial forecasting. This dataset includes:

  • Macroeconomic Event Scripts: Six types of salient macroeconomic announcements (FOMC Minutes, Unemployment Insurance Claims, Employment Situation Reports, GDP Advance Releases, CPI Reports, and PPI Reports) spanning from 2008 to April 2024.
  • High-Frequency Financial Data: Intra-day, 5-minute interval trading data for key U.S. stock indexes (S&P 500, Dow Industrial, NASDAQ) and Treasury bonds (1-Month and 5-Year).
  • Counterfactual Event Scripts: LLM-generated counterfactual versions of the original event scripts, designed to support causal learning.

This dataset is notable for its inclusion of high-frequency data and causally augmented content, differentiating it from previous benchmarks.

Key Insights and Ideas

  • Causality is King: The most significant insight is the critical importance of explicitly modeling causal relationships between macroeconomic events and market movements. Simply correlating text sentiment with price changes is insufficient; understanding why an event impacts the market is key to robust forecasting.
  • LLMs as Causal Augmenters: Large Language Models are not just for text generation; they can be powerful tools for generating synthetic data that helps models learn causal effects by exploring "what-if" scenarios.
  • Multi-Modality Synergy: Combining textual and time-series data is significantly more effective than using either alone, especially when fused intelligently and guided by causal principles.
  • Event Granularity Matters: Not all macroeconomic events are created equal. FOMC Minutes and unemployment-related events appear to have a more pronounced and consistent impact on market movements compared to others like CPI or PPI reports.

Key Results

CAMEF demonstrated superior performance across various financial assets and forecasting horizons.

  • Overall Accuracy: CAMEF outperformed baselines in 24 out of 30 settings, achieving first-place rankings. It showed an average MSE reduction of 62.55% compared to the best single-modality model (iTransformer) and 33.55% compared to the best multi-modality model (TEST).
  • Asset Performance: CAMEF achieved top performance for stock market indices (SPX, INDU, NDX) and the US 1-Month Treasury Bond (USGG1M). Its performance on the US 5-Year Treasury Bond was slightly lower, suggesting these longer-term instruments might be less sensitive to immediate event-driven factors.
  • Component Effectiveness: Ablation studies confirmed the importance of each CAMEF component, with the full model achieving the best results. Causal learning, feature fusion, and the GPT-2 decoder all contributed significantly to the model's performance.
  • Event Impact: FOMC Minutes were found to have the lowest MSE and MAE, indicating their highest predictive power for the S&P 500 index.

Limitations and Future Work

  • Limitations: The paper acknowledges that while CAMEF shows robust performance, the slightly lower accuracy on longer-term Treasury bonds (USGG5YR) might indicate that these assets are influenced by factors beyond immediate event-driven signals. The current LLM prompts for counterfactual generation are designed to maintain structural consistency, but further refinement could explore more complex semantic modifications.
  • Future Work: The authors propose several avenues for future research:
  • Leveraging more advanced LLMs for deeper semantic understanding in textual encoding.
  • Refining cross-modality causal inference mechanisms.
  • Expanding the dataset to include other event types, such as political announcements and corporate news.
  • Further exploring the impact of different post-regressor depths and counterfactual sample sizes on performance.

Summary updated

Abstract

Accurately forecasting the impact of macroeconomic events is critical for investors and policymakers. Salient events like monetary policy decisions and employment reports often trigger market movements by shaping expectations of economic growth and risk, thereby establishing causal relationships between events and market behavior. Existing forecasting methods typically focus either on textual analysis or time-series modeling, but fail to capture the multi-modal nature of financial markets and the causal relationship between events and price movements. To address these gaps, we propose CAMEF (Causal-Augmented Multi-Modality Event-Driven Financial Forecasting), a multi-modality framework that effectively integrates textual and time-series data with a causal learning mechanism and an LLM-based counterfactual event augmentation technique for causal-enhanced financial forecasting. Our contributions include: (1) a multi-modal framework that captures causal relationships between policy texts and historical price data; (2) a new financial dataset with six types of macroeconomic releases from 2008 to April 2024, and high-frequency real trading data for five key U.S. financial assets; and (3) an LLM-based counterfactual event augmentation strategy. We compare CAMEF to state-of-the-art transformer-based time-series and multi-modal baselines, and perform ablation studies to validate the effectiveness of the causal learning mechanism and event types.
BibTeX
@misc{zhang2025camef,
  author = {Yang Zhang and Wenhan Yang and Jun Wang and Qiang Ma and Jie Xiong},
  title = {{CAMEF: Causal-Augmented Multi-Modality Event-Driven Financial Forecasting by Integrating Time Series Patterns and Salient Macroeconomic Announcements}},
  year = {2025},
  eprint = {2502.04592},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.04592},
}
Figure