Scholarlink

<p>With the development of financial temporal data prediction, various derived factors emerge endlessly. Researchers have proposed various derived factors by analyzing the volume and price data, constituting the temporal data of homologous high-dimensional low-rank features. While derived factors can reveal new patterns, they also introduce redundant features and noise. The dimensional disaster caused by the continuous stacking of derived factors not only increases the error, prolongs the calculation time, but also affects the accuracy of the model. To address this problem, this paper proposes a homologous temporal data decomposition and fusion module, which first decomposes the factors by quantifying their information content, then uses tree-based fusion to eliminate redundant information and inter-factor dependence, and finally uses Temporal Convolutional Network (TCN) to perform channel independence temporal analysis. This paper conducts incremental factor robustness and backtesting experiments on multiple stock datasets. The results show that compared with sota methods, the proposed method reduces Mean Squared Error (MSE) by 3.9%, increases Sharpe ratio by 25.2%, and improves average Rank Information Coefficient(RankIC) by 34.1%.</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

A Novel Data Fusion Method for Multi-Dimensional Temporal Data Forecasting of Financial Homologous

  • 期刊论文

  • 发表日期：

    2025年11月25日

查看全文
  • Cheng Zhao,
  • Chengtao Huang,
  • Shuyi Yang,
  • Jie Zhou

摘要

With the development of financial temporal data prediction, various derived factors emerge endlessly. Researchers have proposed various derived factors by analyzing the volume and price data, constituting the temporal data of homologous high-dimensional low-rank features. While derived factors can reveal new patterns, they also introduce redundant features and noise. The dimensional disaster caused by the continuous stacking of derived factors not only increases the error, prolongs the calculation time, but also affects the accuracy of the model. To address this problem, this paper proposes a homologous temporal data decomposition and fusion module, which first decomposes the factors by quantifying their information content, then uses tree-based fusion to eliminate redundant information and inter-factor dependence, and finally uses Temporal Convolutional Network (TCN) to perform channel independence temporal analysis. This paper conducts incremental factor robustness and backtesting experiments on multiple stock datasets. The results show that compared with sota methods, the proposed method reduces Mean Squared Error (MSE) by 3.9%, increases Sharpe ratio by 25.2%, and improves average Rank Information Coefficient(RankIC) by 34.1%.