Scholarlink

<p>Accurately forecasting exchange rate time series is crucial for effective risk management. The longer the forecast time window and the higher the frequency, the more valuable it is for managers to make timely decisions. In light of the task characteristics of exchange rate forecasting, this paper proposes an improved transformer-based model for addressing long sequence time series forecasting problems. Specifically, we introduce a position embedding method for events, which integrates the traditional coding context vector to perform joint location representation learning. Additionally, we explore a periodic enhancement method of features to jointly enhance the transformer-like model architecture’s ability to adaptively learn the long-term and hidden correlations of long sequence time series data. Extensive experimental results on the prediction of four exchange rate data for pairs of opening, high, low, and closing prices at a frequency of 5&#xa0;min show that our model outperforms the prediction of the baseline method.</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

Forecasting of exchange rate time series based on event-aware transformer mode

  • 期刊论文

  • 发表日期：

    2025年3月1日

查看全文
  • Siyi Zhang,
  • Tong Che,
  • Zhiliang Zhu,
  • Guoli Luo,
  • Ping Feng

摘要

Accurately forecasting exchange rate time series is crucial for effective risk management. The longer the forecast time window and the higher the frequency, the more valuable it is for managers to make timely decisions. In light of the task characteristics of exchange rate forecasting, this paper proposes an improved transformer-based model for addressing long sequence time series forecasting problems. Specifically, we introduce a position embedding method for events, which integrates the traditional coding context vector to perform joint location representation learning. Additionally, we explore a periodic enhancement method of features to jointly enhance the transformer-like model architecture’s ability to adaptively learn the long-term and hidden correlations of long sequence time series data. Extensive experimental results on the prediction of four exchange rate data for pairs of opening, high, low, and closing prices at a frequency of 5 min show that our model outperforms the prediction of the baseline method.