Scholarlink

<p>Forecasting exchange rates plays a pivotal role globally in the financial sphere. It heavily influences international trade, cross-border capital flow, and policy formulation. Traditional chronological models struggle with exchange rate data. This is due to the inherent nonlinearity, non-static characteristics, and long-term dependencies in such data. As a result, these models often fail to achieve adequate predictive accuracy and stability. Recently, the impressive effectiveness of long short-term memory networks (LSTM) in predicting temporal trends has gained notable attention. However, their generalization skills are limited in cases where there is scant data or insufficient training. To address these issues, this study introduces Forex-Net, a sophisticated model combining LSTM, transfer learning, and a sophisticated attention mechanism called LinearUnit (LU). This combination is designed to enhance the accuracy and consistency of exchange rate predictions. By incorporating both transfer learning and the LU mechanism into the LSTM model, Forex-Net better handles dynamic features in small datasets. This significantly boosts its generalizability. Experiments demonstrate that Forex-Net outperforms traditional methods such as ARIMA, support vector regression (SVR), and standalone LSTM models. Forex-Net excels at handling variable rate series. It shows a unique ability to predict exchange rates accurately. This approach infuses adaptability into exchange rate series analysis and enhances time series analytical applications. The paper also discusses future directions for further improving the model.</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

Forex-Net: A Hybrid Model for Improved Exchange Rate Prediction Using LSTM and Transfer Learning

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Juntao Tong

摘要

Forecasting exchange rates plays a pivotal role globally in the financial sphere. It heavily influences international trade, cross-border capital flow, and policy formulation. Traditional chronological models struggle with exchange rate data. This is due to the inherent nonlinearity, non-static characteristics, and long-term dependencies in such data. As a result, these models often fail to achieve adequate predictive accuracy and stability. Recently, the impressive effectiveness of long short-term memory networks (LSTM) in predicting temporal trends has gained notable attention. However, their generalization skills are limited in cases where there is scant data or insufficient training. To address these issues, this study introduces Forex-Net, a sophisticated model combining LSTM, transfer learning, and a sophisticated attention mechanism called LinearUnit (LU). This combination is designed to enhance the accuracy and consistency of exchange rate predictions. By incorporating both transfer learning and the LU mechanism into the LSTM model, Forex-Net better handles dynamic features in small datasets. This significantly boosts its generalizability. Experiments demonstrate that Forex-Net outperforms traditional methods such as ARIMA, support vector regression (SVR), and standalone LSTM models. Forex-Net excels at handling variable rate series. It shows a unique ability to predict exchange rates accurately. This approach infuses adaptability into exchange rate series analysis and enhances time series analytical applications. The paper also discusses future directions for further improving the model.