Аннотация
The study examines the problem of intraday forecasting of the EUR/USD currency pair using various neural network architectures, in particular models integrating attention mechanisms. Three neural network architectures were studied: the basic LSTM model, the LSTM model with the Bahdanau attention mechanism, and the Transformer model with the self-attention mechanism. The experiment was conducted on historical minute data for the period from January 2020 to December 2022. The results showed that attentional models are significantly superior to the basic LSTM architecture. The best results were obtained by the Transformer model (MSE=0.185, MAE=0.297, RMSE=0.431, MAPE=7.3%). A detailed analysis confirmed the stability and accuracy of the Transformer model. The identified advantages of attention models justify their prospects for use in algorithmic trading and require further research to optimize and adapt to real trading conditions. In particular, further research may be aimed at integrating attention models with trading strategies and risk management systems, as well as studying their behavior in the face of sudden changes in market volatility. In addition, it is proposed to explore the possibilities of combining attention architectures with other forecasting methods to increase the overall stability and reliability of forecasts in practical trading.