Prediction of Stock Price Time Series using Transformers

This work presents an implementation of the Transformer on the problem of predicting stock prices from time series. The model is compared with ARIMA and a neural network with LSTM cells. We hypothesize that, due to the powerful memory capacity and association between series values, the Transformer would be able to achieve better results than other shallow or deep solutions. The data used in the experiments is the average daily prices of 8 shares of the Ibovespa index in the period of 2008. The obtained results corroborated the hypothesis of superiority of the Transformer which predicted the stock prices with higher accuracy in 60% of the times.