Performance Comparison of Regularized Regression Methods on the Modelling and Forecasting of Bitcoin and Ethereum Prices

Abstract

Over the last decade, investors are interested in model fitting and predicting the future potential value of cryptocurrencies. For this purpose, the multiple linear, Ridge, Lasso and Elastic net regressions allowing for variable selection and regularization are compared. This comparison has yet to be undertaken in the literature. The analysis is implemented using weekly data (from 2015 to 2019) regarding Bitcoin (BTC) and Ethereum (ETH), especially with relation to Google and Wikipedia trends and 17 common factors, including stock market indices, gold and oil prices, central bank interest rates, exchange rates and policy uncertainty. The empirical findings favor the Elastic net approach, which outperforms the others in terms of model fit and predictability. Within the Elastic net framework, while the Google trend for the term "Bitcoin" (positively) has the greatest impact on Bitcoin price, the Chinese Yuan (CNY) to US Dollar (USD) exchange rate (negatively) has the greatest impact on Ethereum price. Based on study findings, essential policy implications are put forward.