Scholarlink

<p>Foreign exchange (Forex) plays a critical role in the global financial market, and accurate forecasting of Forex price movements is essential for investors to manage risk and enhance profitability. However, due to the influence of numerous internal and external factors, Forex time series are inherently volatile and noisy, posing significant challenges for reliable prediction. Existing deep learning approaches often rely on conventional historical financial data and technical indicators, which may not effectively capture the underlying trends in the data. To address this limitation, we propose a novel deep learning model that leverages lag features - differences computed from historical times series values - as input. These features inherently encode trend information and are shown to be statistically stationary, making them more suitable for learning in deep neural networks. The proposed model integrates a 2D Convolutional Neural Network (2D-CNN), a Long Short-term Memory (LSTM) layer, and a dense output layer. The 2D-CNN, equipped with an asymmetric kernel, enables the extraction of features along both time and feature dimensions with different weighting sensitivities. Subsequently, the LSTM layer captures long-term temporal dependencies, while the dense layer produces the final trend prediction. We evaluate our method on four major currency pairs - EUR/USD, CAD/USD, GBP/USD, and JPY/USD. Our method achieves up to 82.3% accuracy for one-day-ahead predictions and up to 64% accuracy for five-day-ahead predictions. Compared to traditional deep learning models such as MLP, 2D-CNN, LSTM, and CNN-LSTM, our approach consistently delivers superior accuracy and faster convergence. These results demonstrate the effectiveness of using lag features and the proposed architecture for robust Forex trend forecasting. Enhanced forecasting accuracy achieved through lag features and deep learning models can offer significant practical benefits, particularly in the areas of trading strategy optimization and 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

A 2D-CNN-LSTM-Based Deep Learning Model for Forex Price Prediction using Lag Features

  • 期刊论文

  • 发表日期：

    2025年9月15日

查看全文
  • Bilguun Narmandakh,
  • Yuming Zhang,
  • Zhen Li,
  • Paul Anderson

摘要

Foreign exchange (Forex) plays a critical role in the global financial market, and accurate forecasting of Forex price movements is essential for investors to manage risk and enhance profitability. However, due to the influence of numerous internal and external factors, Forex time series are inherently volatile and noisy, posing significant challenges for reliable prediction. Existing deep learning approaches often rely on conventional historical financial data and technical indicators, which may not effectively capture the underlying trends in the data. To address this limitation, we propose a novel deep learning model that leverages lag features - differences computed from historical times series values - as input. These features inherently encode trend information and are shown to be statistically stationary, making them more suitable for learning in deep neural networks. The proposed model integrates a 2D Convolutional Neural Network (2D-CNN), a Long Short-term Memory (LSTM) layer, and a dense output layer. The 2D-CNN, equipped with an asymmetric kernel, enables the extraction of features along both time and feature dimensions with different weighting sensitivities. Subsequently, the LSTM layer captures long-term temporal dependencies, while the dense layer produces the final trend prediction. We evaluate our method on four major currency pairs - EUR/USD, CAD/USD, GBP/USD, and JPY/USD. Our method achieves up to 82.3% accuracy for one-day-ahead predictions and up to 64% accuracy for five-day-ahead predictions. Compared to traditional deep learning models such as MLP, 2D-CNN, LSTM, and CNN-LSTM, our approach consistently delivers superior accuracy and faster convergence. These results demonstrate the effectiveness of using lag features and the proposed architecture for robust Forex trend forecasting. Enhanced forecasting accuracy achieved through lag features and deep learning models can offer significant practical benefits, particularly in the areas of trading strategy optimization and risk management.