Cryptocurrency Price forecasting: A Comparative Study of Machine Learning Model in Short-Term Trading

Abstract

In recent years, the expansion of the cryptocurrency market has received significant attention among investors, studies of cryptocurrency price predictions have been conducted in various fields. With the enhancement of machine learning algorithms and increased computational capabilities, machine learning has proved one of the most efficient cryptocurrency prediction methods. However, most studies focused on single digital currency prediction or small-scale algorithm comparison for multiple currencies. This study aims to present a comparative performance of large-scale selected Machine Learning algorithms for cryptocurrency forecasting. Specifically, this paper concentrates on forecasting time series data for a short-term trading period in ten cryptocurrencies (BTC, ETH, ADA, BNB, XRP, DOGE, LUNA, LINK, LTC, and BCH) with ten selected machine learning algorithms (Decision Tree, Linear Regression, Ridge Regression, Lasso Regression, Bayesian Regression, Random Forest, K-Nearest Neighbors, Neural Networks, Gradient Boosting, and Support Vector Machine). Our experiment results show that the Gradient Boosting with the mean square error criterion is superior in predicting most major cryptocurrencies by performing statistical analysis and data visualizations. Additionally, the Random Forest and Decision Tree model built by the Classification and Regression Tree algorithm also shows outstanding performance in certain currencies such as ETH, XRP, LUNA, and LTC. Thus, all three algorithms can help anticipate the short-term evolutions of the cryptocurrency market.