Beyond GARCH in cryptocurrency volatility modelling: superiority of range-based estimators

ABSTRACT

Cryptoassets are extremely volatile with possible volatility jumps and infrastructure noise, making the estimation of true volatility process challenging. When the high-frequency data are not available, the true volatility needs to be estimated to be further studied or forecasted. The GARCH-family models have become a norm in the field. Here, we examine the performance of 6 GARCH-type specifications with 4 distributional assumptions and compare them with 4 non-parametric range-based models built on the daily ‘candles’. Our study focuses on five popular cryptocurrencies (Bitcoin, Ethereum, BNB, XRP, and Dogecoin) between 1 July 2019 and 30 September 2022, utilizing Binance 5-minute data for realized measures as the high-frequency estimators of the true volatility process. The results reveal that the Garman-Klass estimator clearly outperforms the GARCH-family models in all studied settings, and the other range-based estimators remain competitive with the GARCH-family models. These results are crucial for studies on volatility in cryptoassets where using the GARCH-type models is a standard. When the high-frequency data are not available, the range-based estimators, and the Garman-Klass estimator in particular, should be preferred as proxies for the true volatility process over the GARCH-type models, be it in the in-sample, more qualitative studies, or the forecasting, out-of-sample exercises.

KEYWORDS:

• Cryptoasset
• cryptocurrency
• volatility
• GARCH
• Garman-Klass

JEL CLASSIFICATION:

• C32
• C53
• G23

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ We use the terms cryptoassets and cryptocurrencies interchangeably throughout the text.

² Currently freely available at https://data.binance.vision/?prefix=data/spot/daily/klines/.

³ The results for the bipower variation and the realized kernel are available in Tables A1 and Table A2. in the Appendix. The implications are qualitatively parallel to the realized volatility case.

Additional information

Funding

This work was supported by the Cooperatio Program at Charles University, research area Economics. Support from the Czech Science Foundation (project 23-06606S “Deep dive into decentralized fiMarket microstructure, and behavioral and psychological patterns”) is also highly appreciated.