Scholarlink

<p>Accurate volatility forecasting is essential for financial market risk management, however, existing methods mostly rely on a limited set of features and fail to adequately incorporate multimodal external information. Most prior studies rely predominantly on daily frequency data, limiting their ability to represent complex market behaviors embedded in high-frequency intraday data. To overcome these limitations, this paper proposes a time series forecasting method based on the TimeXer model, enhancing multimodal feature integration by incorporating both textual and image data. This enriched approach effectively represents comprehensive market information, improving the model’s capability to characterize volatility patterns. To address the inherent lag associated with traditional time series forecasting and enhance practical applicability, this paper adopts half-daily realized volatility measures and further subdivides the volatility forecasting task into intraday and interday predictions, upon which an adaptive time frequency selective model fusion framework is constructed. Empirical results confirm that the proposed framework demonstrates superior forecasting accuracy and robustness in the Shanghai Stock Exchange (SSE) Index market. Additionally, by calculating Value at Risk (VaR), the study underscores the model’s practical value for financial risk management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multimodal Stock Volatility Forecasting Across Multiple Timeframes Via TimeXer

  • 期刊论文

  • 发表日期：

    2026年4月29日

查看全文
  • Yilun Zhang,
  • Shiji Liu,
  • Yufeng Shi

摘要

Accurate volatility forecasting is essential for financial market risk management, however, existing methods mostly rely on a limited set of features and fail to adequately incorporate multimodal external information. Most prior studies rely predominantly on daily frequency data, limiting their ability to represent complex market behaviors embedded in high-frequency intraday data. To overcome these limitations, this paper proposes a time series forecasting method based on the TimeXer model, enhancing multimodal feature integration by incorporating both textual and image data. This enriched approach effectively represents comprehensive market information, improving the model’s capability to characterize volatility patterns. To address the inherent lag associated with traditional time series forecasting and enhance practical applicability, this paper adopts half-daily realized volatility measures and further subdivides the volatility forecasting task into intraday and interday predictions, upon which an adaptive time frequency selective model fusion framework is constructed. Empirical results confirm that the proposed framework demonstrates superior forecasting accuracy and robustness in the Shanghai Stock Exchange (SSE) Index market. Additionally, by calculating Value at Risk (VaR), the study underscores the model’s practical value for financial risk management.