Performance Comparison of ARIMA and Machine Learning Regression Techniques in Time Series Forecasting of Bitcoin Prices

Abstract

At present, Bitcoin is the most popular and traded cryptocurrency, causing its price to fluctuate on a daily basis. Therefore, research to find a suitable method for forecasting the Bitcoin price has attracted attention. ARIMA and machine learning regression techniques are techniques used for forecasting various situations such as gold price and oil price forecasting. ARIMA is a classical technique while machine learning regression techniques are new techniques such as Multilayer perceptron, Radial basis function and Support vector regression. This research aims to compare the performance of ARIMA and machine learning regression techniques in forecasting Bitcoin price. Both techniques were used to create time series forecasting models for daily Bitcoin price. The data used for creating models were the opening price, the highest price, the lowest price, the closing price, the volume and market cap from www.coinmarketcap.com during 1 January 2017 to 31 December 2019. In performance evaluation of the models, sliding window technique was used to separate data into training and testing sets. Furthermore, mean absolute error (MAE) and root mean square error (RMSE) were also used as criteria to compare performance of the models. From the experiments, it was found that ARIMA showed the best performance with the lowest MAE and RMSE for forecasting the opening price, the highest price, the lowest price, the closing price, the volume and market cap.