Multivariate Financial Time Series Forecasting with Deep Learning.

Today, and aware of how unexpected the events that govern the market trend can be, forecasting financial time series has become a priority for everyone, a field in which computational intelligence with networks par excellence, Long-term and short-term neural networks (LSTM) and Gated Recurrent Unit (GRU), has taken the center of the stage. To avoid long-term dependency problems, given their unique storage unit structure, these networks are postulated as the best option when predicting financial time series. Thus, the motivation of this paper is to compare the transformer model with Long and Short Term Neural Networks (LSTM) and Gated Recurrent Unit (GRU), where the data set of the NASDAQ 100 index will be used, with a time interval "tick by tick" and with a range from January 2020 to July 2020, since those of deep learning (DL) models have presented significantly better results than their traditional counterparts, results with which they will be compared. The results will be categorized according to their performance in a ranking from 1 to 3, where number 1 would be the best option when predicting the behavior of the selected index, demonstrating if the Transformer architecture presents a better performance versus GRU and LSTM.