A comprehensive approach to Bitcoin forecasting using neural networks

Abstract

This paper provides a comprehensive approach to Bitcoin price, returns, direction and volatility forecasting. It compares ARIMA and GARCH models to neural network (NN) autoregression and Jordan NN in their forecasting performances, using internal and external factors. Robustness of the results is verified across bearish, bullish and stable market conditions. The results are not unambiguous considering price, returns or volatility forecasting, when compared using different performance measures or through different periods. Return and volatility forecasting yields to stable results no matter the model or period observed. NNs in general emerge as optimal for return and direction forecasting, ARIMAX and NNARX for price forecasting, while for volatility forecasting all models yield comparable results. Price forecasting yields the best prediction accuracies, while JNNX performed poorly. However, the inclusion of other machine learning methods and/or different variables as well as recent crisis emerged from war circumstances can be seen as limiting factors.