Cryptocurrency Price Forecast with Use of Time Series Model, Conventional Statistic Model and Machine Learning Approaches

Cryptocurrency is an emerging virtual currency gathering attention from the public, which is difficult to counterfeit or double-spend. One of the most significant discussions in cryptocurrency is its extremely high volatility which brings challenges on price prediction. Therefore, lots of research have been conducted to predict cryptocurrency by implementing different models. In this paper, various time series models, conventional linear models, and machine learning models are compared in terms of predictive performance on BTC-USD prices based on historical 5-year daily information and technical features derived. On this basis, various metrics are adopted including AIC, RMSE, MAE, and R-Square and are respectively evaluated and compared. Time series models, ARIMA and GARCH, have relatively poor predictive performance. For statistic models, linear regression, Ridge and Lasso are evaluated respectively, where both regularized models are not able to outperform the linear regression. Regarding to machine learning models, ensemble tree methods including random forest and LightGBM have relatively better performance than other types of models. Among all models tested for the same split, random forest has the lowest error and highest coefficient of determination, and its predictions are the most accurate. These results shed light on choosing from different models in cryptocurrency price prediction.