Price Prediction for Bitcoin: Does Periodicity Matter?

Abstract

Purpose: A major challenge traders, speculators and investors are grappling with is how to accurately forecast Bitcoin price in the cryptocurrency market. This study is aimed to uncover the best model for the forecasts of Bitcoin price as well as to verify the price series that offers the best predictions performance under different periodicity of datasets. Design/methodology/approach: The study adopts three different data periods to verify whether frequency matters in forecasting Bitcoin price. The Bitcoin price, from 01/01/15 to 11/01/2021, is trained and validated on selected forecast models, including the Naïve, Linear, Exponential Smoothing Model, ARIMA, Neural Network, STL and Holt-Winters filters. Five forecast accuracy measures (RSME, MAE, MPE, MAPE and MASE) are applied to confirm the best performing model. The Diebold‐Mariano test is used to compare the forecasts based on the daily price with those based on the weekly and monthly. Findings: Based on the accuracy measures, the results indicate that the Naïve model provides more accurate performance for the daily series, while the linear model outperforms others for the weekly and monthly series. Using the Diebold‐Mariano statistics, there is evidence that forecasting Bitcoin price is not sensitive to the data periodicity. Research limitations/implications: The study has a major limitation, which is the shared sentiment to apply actual Bitcoin price series, and not the returns or log transformation for the forecast models. Notably, actual data may sometimes be loud, hence increasing the possibility of over predictions. Originality/value: In forecasting, different approaches have been used, this paper compares outputs of both statistical and machine learning methods in order to arrive at the best option for the Bitcoin price forecasts. Hence, we investigate whether the machine learning tools offer better forecasts in terms of lower error and higher model’s accuracy relative to the traditional models.