Unlocking the black box of sentiment and cryptocurrency: What, which, why, when and how?

Abstract

We evaluate the impact of a large set of daily sentiment measures for predicting Ethereum (ETH) returns using Machine Learning (ML) methods. We examine ETH predictability and evaluate 5 W's: What, Which, When, Why, and hoW. What ML methods work best? Which variables robustly predict ETH returns? When and why does predictability occur? And how can we improve predictability? We extract information from fifty sentiment measures from Refinitiv's MarketPsych Analytics using ML methods including Lasso, Elastic Net, Principal Components, Partial Least Squares, Neural Net and Random Forest. We then apply an ensemble procedure that exponentially weights forecasts from these traditional ML methods based on recent MSFE criteria. By discounting past model performance, our ensemble procedure accommodates time variation in model selection and generates investment gains and significant out-of-sample pre- dictability. Our study offers practical implications for investing in ETH, including considering an array of sentiment measures, diversifying your model forecasts using an ensemble approach, and the importance of transaction costs in trading simulations.