Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fusing Narrative Semantics for Financial Volatility Forecasting

Yaxuan Kong, Yoontae Hwang, Marcus Kaiser, Chris Vryonides, Roel Oomen, Stefan Zohren

TL;DR

The paper tackles volatility forecasting by fusing structured time-series features with temporally faithful news embeddings derived from TiMaGPT, addressing two core challenges: modality fusion and look-ahead bias. It introduces M2VN, a three-branch architecture that encodes price signals, news narratives, and calendar cues into a shared latent space, uses spectral filtering to capture persistent cycles, and applies an InfoNCE-based alignment loss to promote cross-modal coherence. Empirical results across seven large-cap U.S. stocks show M2VN consistently beats classical HAR-based models and modern DL baselines, with ablations demonstrating that both trading volume and news embeddings provide complementary predictive signals. The work highlights the practical value of integrating narrative semantics with market data for improved volatility forecasting and risk management in dynamic markets.

Abstract

We introduce M2VN: Multi-Modal Volatility Network, a novel deep learning-based framework for financial volatility forecasting that unifies time series features with unstructured news data. M2VN leverages the representational power of deep neural networks to address two key challenges in this domain: (i) aligning and fusing heterogeneous data modalities, numerical financial data and textual information, and (ii) mitigating look-ahead bias that can undermine the validity of financial models. To achieve this, M2VN combines open-source market features with news embeddings generated by Time Machine GPT, a recently introduced point-in-time LLM, ensuring temporal integrity. An auxiliary alignment loss is introduced to enhance the integration of structured and unstructured data within the deep learning architecture. Extensive experiments demonstrate that M2VN consistently outperforms existing baselines, underscoring its practical value for risk management and financial decision-making in dynamic markets.

Fusing Narrative Semantics for Financial Volatility Forecasting

TL;DR

The paper tackles volatility forecasting by fusing structured time-series features with temporally faithful news embeddings derived from TiMaGPT, addressing two core challenges: modality fusion and look-ahead bias. It introduces M2VN, a three-branch architecture that encodes price signals, news narratives, and calendar cues into a shared latent space, uses spectral filtering to capture persistent cycles, and applies an InfoNCE-based alignment loss to promote cross-modal coherence. Empirical results across seven large-cap U.S. stocks show M2VN consistently beats classical HAR-based models and modern DL baselines, with ablations demonstrating that both trading volume and news embeddings provide complementary predictive signals. The work highlights the practical value of integrating narrative semantics with market data for improved volatility forecasting and risk management in dynamic markets.

Abstract

We introduce M2VN: Multi-Modal Volatility Network, a novel deep learning-based framework for financial volatility forecasting that unifies time series features with unstructured news data. M2VN leverages the representational power of deep neural networks to address two key challenges in this domain: (i) aligning and fusing heterogeneous data modalities, numerical financial data and textual information, and (ii) mitigating look-ahead bias that can undermine the validity of financial models. To achieve this, M2VN combines open-source market features with news embeddings generated by Time Machine GPT, a recently introduced point-in-time LLM, ensuring temporal integrity. An auxiliary alignment loss is introduced to enhance the integration of structured and unstructured data within the deep learning architecture. Extensive experiments demonstrate that M2VN consistently outperforms existing baselines, underscoring its practical value for risk management and financial decision-making in dynamic markets.
Paper Structure (14 sections, 1 theorem, 15 equations, 3 figures, 2 tables)

This paper contains 14 sections, 1 theorem, 15 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Financial Volatility Forecasting
  4. Text-Enhanced Time Series Forecasting
  5. Methodology
  6. Problem Formulation
  7. M2VN: Multi-Modal Volatility Network
  8. Experiments
  9. Implementation Details
  10. M2VN is Useful for Volatility Forecasting?
  11. Is Volume a Proxy for News Information?
  12. Is News Embedding Predictive?
  13. Showcasing Forecast Quality
  14. Conclusion

Key Result

proposition 1

Let $\mathbf h=(h_0,\dots,h_{T-1})\in(\mathbb C^{d})^{T}$ and denote by $\widehat{\mathbf h}=\mathcal{F}\mathbf h$ its unitary discrete Fourier transform For $1 \le k \le T$ let $I_k\subseteq\{0,\dots,T-1\}$ collect the $k$ indices of largest energy $\|\widehat{\mathbf h}_{\omega}\|_2$ (ties broken arbitrarily) and set Then $\mathbf h_k$ minimises the approximation error where $\|\cdot\|_F$ is

Figures (3)

  • Figure 1: Illustration of two challenges: data fusion across modalities and mitigation of look-ahead bias.
  • Figure 2: The architecture of M2VN, a multi-modal volatility network for financial volatility forecasting.
  • Figure 3: Visual comparison of out-of-sample forecast quality for the KO (top) and CMCSA (bottom) datasets over the test period of 2021-2023 . The M2VN model's predictions (green) are plotted against the ground truth realized volatility (black), alongside forecasts from the HAR-X (Lasso) (orange) and TimeXer (blue) baselines. The M2VN demonstrates a superior ability to track the fluctuations and sharp peaks of the ground truth data compared to the other models.

Theorems & Definitions (1)

  • proposition 1: Best $k$-term Fourier approximation