Go to ICLR 2025 Conference homepageTransformers versus LSTMs for electronic trading
The rapid advancement of artificial intelligence has seen widespread application of long short-term memory (LSTM), a type of recurrent neural network (RNN), in time series forecasting. Despite the success of Transformers in natural language processing (NLP), which prompted interest in their efficacy for time series prediction, their application in financial time series forecasting is less explored compared to the dominant LSTM models. This study investigates whether Transformer-based models can outperform LSTMs in financial time series forecasting. It involves a comparative analysis of various LSTM-based and Transformer-based models on multiple financial prediction tasks using high-frequency limit order book data. A novel LSTM-based model named DLSTM is introduced alongside a newly designed Transformer-based model tailored for financial predictions. The findings indicate that Transformer-based models exhibit only a marginal advantage in predicting absolute price sequences, whereas LSTM-based models demonstrate superior and more consistent performance in predicting differential sequences such as price differences and movements.