Predição da Cotação Real/Bitcoin usando a Rede Neural Long Short Term Memory

Abstract

Cryptocurrencies have become an important financial asset for com- panies, governments and civilians, being accepted for financial transactions in place of traditional currencies. However, cryptocurrencies, which are attractive for some reasons such as not having government control, have high volatility, making price predictions important in order to have a basis for deciding when to buy and when to sell this type of asset, in order to maximize profits. As pre- diction by people demands a specialist, which is expensive and also subject to biases and errors, several studies have investigated the application of artificial intelligence computational techniques for prediction. This study investigates the application of an LSTM neural network for predicting the price of the Bitcoin cryptocurrency in Reais, analyzing parameters of the neural network, together with the network modeling for the prediction of a single value for a given future time horizon and the prediction of multiple values for multiple future instants of time. The results obtained indicate that the single output model has a good performance for short-term loss, mainly in the prediction for the next day (horizon 1) with a mean absolute percentage error (MAPE) of 2.57% in its best configuration, however this error is not linearly scalable when the prediction horizon is increased. The multiple-output model produced higher MAPE values for short-term predictions (horizons 1, 3 and 5), however the error increase rate is lower than that of the simple horizon model.