Prediction of Bitcoin price based on manipulating distribution strategy

Abstract Since Bitcoin has been the most popular digital currency in the global financial market, the prediction of its price has been an important area in finance. Recently, numerous researches on prediction of financial indices including Bitcoin based on machine learning techniques have been developed. However, little studies pay attention to the strategy of manipulating original distribution to obtain better performances. Because the return data of Bitcoin prices are highly concentrated near zero, small changes of values in the components can cause different results. Based on this characteristic, we propose flattening distribution strategy (FDS) based on the copula theory as a strategy of the manipulating distribution of components artificially to improve the prediction of Bitcoin price return. We consider multilayer perceptron (MLP), recurrent neural networks (RNN), and long short-term memory (LSTM) to assess the performances of FDS. Finally, we find that the proposed algorithms based on FDS improve significantly the prediction accuracy of the return of Bitcoin price for each of the three architectures.