Forecasting with Competing Models of Daily Bitcoin Price in R

Abstract

Bitcoin price exhibits patterns predictable on its historical pasts. We adopt ARIMA(auto), ARIMA(fix)models and the Holt-Winters filter (HWF) with trend plus additive seasonal HWF (𝛾[0,1]), and no seasonality HWF (𝛾[False]) to forecast the price of Bitcoin under three datasets–Actual (observed), Polynomial (fitted) and STL-Trend (fitted). We apply daily time-series from 1/09/2014–28/12/2020,and establish 18 models to forecast the price of Bitcoin. The results show that HWF (𝛾[0,1]) with lower limit fitted on STL-Trend provides the best prediction on the first training-sample, while ARIMA(fix) fitted on actual data outperform in the second training-set with the smallest Mean-Absolute-Error (MAE). The training-set forecast performance of the ARIMA(fix) for the actual function provides better performance with the least MAE. The HWF is appropriate for prediction of the daily Bitcoin price with the generalise STL-Trend function, but ARIMA(fix) is more accurate for the actual series.

Keywords

• bitcoin
• arima
• holt-winters
• mean absolute error