Comparison of CNN-Transformer and LSTM models for forex market forecasting
Loading...
Supplementary material
Other Title
Authors
Varastehpour, S.
Abdolahi, A.
Modares, A.F.A.
Varastehpour, Soheil
Abdolahi, A.
Modares, A.F.A.
Varastehpour, Soheil
Author ORCID Profiles (clickable)
Degree
Grantor
Date
2025
Supervisors
Type
Conference Contribution - Paper in Published Proceedings
Ngā Upoko Tukutuku (Māori subject headings)
Keyword
stock price analysis
prediction
Convolutional Neural Network (CNN)
Long Short-Term Memory (LSTM)
deep-learning algorithms
computer modelling
algorithms
prediction
Convolutional Neural Network (CNN)
Long Short-Term Memory (LSTM)
deep-learning algorithms
computer modelling
algorithms
ANZSRC Field of Research Code (2020)
Citation
Varastehpour, S., Abdolahi, A., Modares, A. F. A., & Varastehpour, S. (2025). Comparison of CNN-Transformer and LSTM models for forex market forecasting In S. Varastehpour & M. Shakiba (Eds.), Proceedings: AIOT Global Summit 2025: Economic Growth, 15–16 July (pp. 73–79). ePress, Unitec. https://doi.org/10.3.4074/proc.250114
Abstract
The foreign exchange (Forex) market is one of the largest and most volatile financial markets worldwide, making accurate trend prediction a significant challenge. In this study, we compare the performance of a traditional long short-term memory (LSTM) model with a hybrid convolutional neural network-transformer (CNN-transformer) model for forecasting Forex trends. Using a 20-year dataset covering major currency pairs (AUD/USD, CAD/CHF, EUR/ GBP, GBP/USD), we evaluate both models based on common metrics such as mean square error (MSE), root mean square error (RMSE), relative mean error (RME), and mean absolute percentage error (MAPE). The results show that the LSTM model generally outperforms our proposed hybrid CNN-transformer model in most cases. Based on the findings of this study, it can be concluded that our proposed hybrid model performed poorly for some currency pairs, such as AUD/USD, CAD/CHF and EUR/GBP, while it showed better performance than the baseline model for the GBP/USD pair. Therefore, it can be inferred that our proposed model works better for specific data that is less noisy and less volatile.
Publisher
Unitec ePress
Permanent link
Link to ePress publication
DOI
https://doi.org/10.34074/proc.250114
Copyright holder
Authors
Copyright notice
CC BY-NC Attribution-NonCommercial 4.0 International