Characteristics of Bitcoin Volatility at High Frequency

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
School of Business | Master's thesis
Ask about the availability of the thesis by sending email to the Aalto University Learning Centre
Degree programme
This thesis models and examines the realized bitcoin volatility calculated from five-minute intraday squared returns and its unconditional distribution. The study also explores the intraday periodicity and volatility persistence in bitcoin markets. In addition, the thesis tests if there are day-of-the-week effects for bitcoin returns and volatility. Lastly, the study compares model-free realized volatility measures to linear and nonlinear GARCH-models and tests their prediction power to find the most efficient model for bitcoin volatility. I use all transactions for BTCUSD between January 2014 and February 2018, using data from four cryptocurrency markets which CME uses to calculate their bitcoin price index, which represent currently the closest thing to an official BTCUSD price. This amounts to a total of 67,600,230 trades. After the descriptive analysis of the data, I calculate five-minute interval logarithmic returns from the aggregate dataset. I use kernel density functions to create the volatility distributions, and I employ various tests and regressions to research the intraday periodicity and persistence of bitcoin volatility. I construct 10 different GARCH specifications and compare their fit in-sample and out-of-sample by using four different information criteria, two robust loss functions and adjusted coefficient of determination as the comparison measures. The robustness of the results is confirmed using different intraday sampling intervals, namely 1-minute and 30-minute intervals The main findings of this thesis are that the distribution for realized bitcoin volatility are extremely fat tailed, or leptokurtic. Furthermore, the volatility has a long lasting and high persistence, and the autocorrelations are statistically significant for long periods. The findings for unconditional distributions, and volatility persistence are consistent and similar with prior research on currencies and equities. Interestingly and on the contrary to currencies and equities, bitcoin intraday periodicity does not show the typical U-shaped pattern, which is likely due to absence of institutional traders and market makers, and the possibility to trade 24 hours a day. In addition, nonlinear GARCH-models, especially CGARCH gives the best in-sample model fit for bitcoin, which also performs best with out-of-sample forecasts, suggesting the importance of having both short-run and long-run components of conditional variance when modeling bitcoin volatility using intraday data. This study adds to the existing literature by suggesting a transfer of academic interest from the traditional asset classes to cryptocurrencies and how to model and estimate their unconditional distributions and volatility. According to my best knowledge, this is the first study to model the unconditional distribution of the realized volatility for bitcoin, and the first to compare RV and GARCH models for bitcoin using intraday data.
Thesis advisor
Suominen, Matti
volatility, realized volatility, GARCH, bitcoin, cryptocurrency, forecasting