Volatility regime analysis of Bitcoin price dynamics using Markov-switching GARCH models

Abstract

One of many applications of machine learning is that of correctly classifying market and economic phases. A popular approach to this is using Markov switching models. This paper will look at the implementation of Markov switching generalised autoregressive conditional heteroscedastic (MSGARCH) models with nor-mal and t-distributed innovations to recent Bitcoin/US Dollar price dynamics, and also show that these can be an improvement over single-regime models of the same kind, by demarcating high and low volatility regimes. Furthermore, we also look at both maximum likelihood estimation, and Bayesian estimation via Markov Chain Monte Carlo. The predictive performance is also analysed using risk management tools such as value-at-risk and expected shortfall, and we show that the two-regime model with t-distributed innovations provides the best fit.