Uso de algoritmos para predição de séries temporais aplicado ao mercado de criptomoedas

Abstract

The data processing capacity fast growth resulting from the evolution of computing power, attached with machine learning and artificial intelligence techniques are able to perform accurate time series predictions. This work addresses the concepts of Artificial Intelligence, Machine Learning, Cryptocurrencies and mainly Time Series calculation models, such as: Facebook’s Prophet, ARIMA and LSTM. This work aims to verify the ability of these models to predict the value of cryptocurrencies. A theoretical framework is presented and its main references used, is also presented the main technologies that are used in the current days, as the correlated works. The performance metrics adopted to evaluate the models (MAPE and RMSE) and the tests performed using the tools Facebook’s Prophet, ARIMA and LSTM are presented. The models developed were able to follow the up and down trends of the time series. Facebook’s Prophet scored 10.93%, 16.32%, and 40.14% mean error percentage for 1-day, 1-week, and 1-month forecasts respectively. The ARIMA model, on the other hand, obtained 3.41%, 13.60% and 17.98% of average error percentage for 1-day, 1-week and 1-month forecasts. The LSTM model obtained 8.00%, 7.46% and 23.37% of average error percentage for 1-day, 1-week and 1-month forecasts. Finaly, this study present the final considerations and future works.