Forecasting of Cryptocurrency Prices Using Machine Learning

Abstract

Our study is devoted to the problems of the short-term forecasting cryptocurrency time series using machine learning (ML) approach. Focus on studying of the financial time series allows to analyze the methodological principles, including the advantages and disadvantages of using ML algorithms. The 90-day time horizon of the dynamics of the three most capitalized cryptocurrencies (Bitcoin, Ethereum, Ripple) was estimated using the Binary Autoregressive Tree model (BART), Neural Networks (multilayer perceptron, MLP) and an ensemble of Classification and Regression Trees models—Random Forest (RF). The advantange of the developed models is that their application does not impose rigid restrictions on the statistical properties of the studied cryptocurrencies time series, with only the past values of the target variable being used as predictors. Comparative analysis of the predictive ability of the constructed models showed that all the models adequately describe the dynamics of the cryptocurrencies with the mean absolute persentage error (MAPE) for the BART andMLPmodels averaging 3.5%, and for RFmodels within 5%. Since for trading perspective it is of interest to predict the direction of a change in price or trend, rather than its numerical value, the practical application of BART model was also demonstrated in the forecasting of the direction of change in price for a 90-day period. To this end, a model of binary classification was used in the methodology for assessing the degree of attractiveness of cryptocurrencies as an innovative financial instrument. Conducted computer simulations have confirmed the feasibility of using the machine learning methods and models for the short-term forecasting of financial time series. Constructed models and their ensembles can be the basis for the algorithms for automated trading systems for Internet trading.