improving bitcoin price prediction power by time-scale decomposition and gmdh-type neural network: a comparison of different periods and features

This paper aims to improve the predictability power of a machine learning method by proposing a two-stage prediction method. we use group modeling data handling (gmdh)-type neural network method to eliminate the user role in feature selection. to consider recent shocks in bitcoin market, we consider three periods, before covid-19, after covid-19, and after elon musk's tweeter activity. using time-scale analysis, we decomposed the data into different scales. we further investigate the forecasting accuracy across different frequencies. the findings show that in shorter period the first, second and third lag of daily prices and trade volume produce valuable information to predict bitcoin price while the seven days lag can improve the prediction power over longer period. the results indicate a better performance of the wavelet base gmdh-neural network in comparison with the standard method. this reveals the importance of trade frequencies' impact on the forecasting power of models.