QUANTIFYING CRYPTOCURRENCY UNCERTAINTY WITH NATURAL LANGUAGE PROCESSING FOR PRICE RETURNS FORECASTING.

Abstract

I quantify uncertainty in the cryptocurrency market by using Lucey et al.’s (2021) lexical methodology to construct indices pertaining to cryptocurrency price and policy uncertainty. I extend Lucey et al.’s (2021) methodology by using alternative social media forum data over traditional news media texts. Additionally, I apply unsupervised topic modelling and a novel supervised hedge detection approach to construct social media based uncertainty indices with significantly less data and resources. I then evaluate their added predictive value by using them as predictors in Bitcoin price returns and directional price returns forecasting. I find mixed results for out of sample performance, where forecasting models using the social media based uncertainty indices marginally outperform in shorter horizons but underperform in longer ones as compared to Lucey et al.’s (2021) news based uncertainty indices. However, the directional price returns forecast shows promising results having outperformed benchmarks from similar literature on cryptocurrency price movement prediction.