Paper: arXiv 2510.20699

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

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.

Complexity vs Empirical Score

  • Math Complexity: 4.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Street Traders — empirical and applied, lower math complexity

Why this score: The paper utilizes deep learning and specialized loss functions, indicating moderate mathematical complexity, but places a strong emphasis on data handling, look-ahead bias mitigation, and backtesting against baselines, making it highly implementation-heavy.

Research Flowchart

  flowchart TD
  A["Research Goal: Forecast Financial Volatility using<br>Unstructured News + Structured Market Data"]
  B["Inputs: Price Series &<br>Point-in-Time News (Time Machine GPT)"]
  C["Methodology: M2VN<br>Multi-Modal Volatility Network"]
  D["Key Innovation:<br>Auxiliary Alignment Loss"]
  E["Computation:<br>Deep Learning Fusion & Alignment"]
  F["Outcome: Consistent Outperformance<br>Mitigates Look-ahead Bias"]

  A --> B
  B --> C
  C --> D
  D --> E
  E --> F

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes M2VN, a Multi-Modal Volatility Network that aligns and fuses news embeddings with time series data for volatility forecasting.

Source sentence
We propose M2VN: Multi-Modal Volatility Network, a novel framework to align and fuse news embeddings with time series data for the task of financial volatility forecasting.

Assumes 2

  • News articles are temporally aligned on a daily basis. data
    Source sentence
    All news articles are temporally aligned on a daily basis in a structured format.
  • The dataset consists of daily end‑of‑day OHLCV and full‑text news for seven large‑cap U.S. equities. market
    Source sentence
    ... extract both end-of-day price information (open, high, low, close and volume) and the full text of associated news articles from FNSPID [14].

Data

  • Seven large‑cap U.S. equities panel · FNSPID market-prices public
    2013-01-01 to 2023-12-31 · KO, CMCSA, COP, GILD, MRK, NKE, ORCL
    Source sentence
    ... extract both end-of-day price information (open, high, low, close and volume) and the full text of associated news articles from FNSPID [14].
  • M2VN pre‑processed dataset and code · GitHub repository benchmark public
    Source sentence
    The pre‑processed dataset, raw news article data, and implementation code are publicly accessible at https://github.com/Yoontae6719/M2VN-Multi-Modal-Learning-Network-for-Volatility-Forecasting

Recognised sources: FNSPID

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • M2VN consistently outperforms existing baselines on volatility forecasting. performance
    Source sentence
    Extensive experiments demonstrate that M2VN consistently outperforms existing baselines, underscoring its practical value for risk management and financial decision‑making.
  • Ablating the cross‑modal alignment loss reduces mean QLike by 11.7% and MAPE by 3.6%. predictive-accuracy
    Source sentence
    ablation of our backbone that omits the cross-modal alignment loss Lalign, mean QLike and MAPE fall by 11.7 % and 3.6 %, respectively.
  • Including trading volume improves forecasting accuracy, reducing average MAPE by 2.86% and QLike by 5.64% across all stocks. predictive-accuracy
    Source sentence
    On average, incorporating volume reduces MAPE by 2.86±3.06% and QLike by 5.64±2.40% across all stocks.
  • News embeddings improve performance; e.g., for Coca‑Cola, QLike drops from 0.0641 to 0.0589 and MAPE from 0.2904 to 0.2771. predictive-accuracy
    Source sentence
    Coca-Cola (KO), the QLike decreased from 0.0641 to 0.0589 and MAPE from 0.2904 to 0.2771 when news was incorporated.
  • Visual comparison shows M2VN tracks ground‑truth volatility fluctuations and sharp peaks better than HAR‑X and TimeXer baselines. performance
    Source sentence
    The M2VN demonstrates a superior ability to track the fluctuations and sharp peaks of the ground truth data compared to the other models.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 53 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 3 papers in the archive (2 by independent authors)

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (46)

  1. Time-mmd: Multi-domain multimodal dataset for time series analysis (2024)
  2. Reversible instance normalization for accurate time-series forecasting against distribution shift (2021)
  3. Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
  4. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting (2021)
  5. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis (2023)
  6. Representation learning with contrastive predictive coding (2018) — link
  7. A wavelet tour of signal processing (1999)
  8. Harvesting the HAR-X Volatility Model (2024)
and 38 more
  1. On the estimation of security price volatilities from historical data (1980)
  2. Estimating variance from high, low and closing prices (1991)
  3. The extreme value method for estimating the variance of the rate of return (1980)
  4. Time machine GPT (2024) — link
  5. Are transformers effective for time series forecasting? (2023)
  6. Are language models actually useful for time series forecasting? (2024)
  7. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (2023) — link
  8. Timexer: Empowering transformers for time series forecasting with exogenous variables (2024)
  9. Fnspid: A comprehensive financial news dataset in time series — → 2402.06698
  10. One fits all: Power general time series analysis by pretrained lm (2023)
  11. Large language models are zero-shot time series forecasters (2023)
  12. Time-llm: Time series forecasting by reprogramming large language models (2023) — link
  13. Lookahead bias in pretrained language models (2024)
  14. Financial Statement Analysis with Large Language Models — → 2407.17866
  15. Forecasting oil price realized volatility using information channels from other asset classes (2017)
  16. Chronos: Learning the language of time series (2024) — link
  17. Decision-informed Neural Networks with Large Language Model Integration for Portfolio Optimization — → 2502.00828
  18. The impact of sentiment and attention measures on stock market volatility (2020)
  19. Forecasting daily variability of the S&P 100 stock index using historical, realised and implied volatility measurements (2005)
  20. Modelling the persistence of conditional variances (1986)
  21. Sentiment correlation in financial news networks and associated market movements (2021)
  22. Answering the skeptics: Yes, standard volatility models do provide accurate forecasts (1998)
  23. A simple approximate long-memory model of realized volatility (2009)
  24. Compressed Sensing and Best \emph k -Term Approximation (2006) — link
  25. Introduction to compressed sensing (2012)
  26. Compressed sensing and best k-term approximation (2009)
  27. Forecasting volatility in financial markets: A review (2003)
  28. Volatility and correlation forecasting (2006)
  29. Market risk analysis, practical financial econometrics (2008)
  30. Realised volatility forecasting: Machine learning via financial word embedding (2021) — link
  31. Textual Regression for Realized Volatility: A Model for Long-Term Forecasting (2025)
  32. Financial news predicts stock market volatility better than close price (2018)
  33. Stock market volatility forecasting: Do we need high-frequency data? (2021)
  34. Beyond the S&P 500: examining the role of external volatilities in market forecasting (2024)
  35. A practical guide to harnessing the har volatility model (2021)
  36. A machine learning approach to volatility forecasting (2023)
  37. Large Language Models for Financial and Investment Management: Applications and Benchmarks. (2024)
  38. Large Language Models for Financial and Investment Management: Models, Opportunities, and Challenges. (2024)

3 of 46 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.