Sequential Learning of Cryptocurrency Volatility Dynamics: Evidence Based on a Stochastic Volatility Model with Jumps in Returns and Volatility

Jing‐Zhi Huang, Zhijian Huang, Li XuQuarterly 2021 · DOI

Abstract

This paper studies the dynamics of cryptocurrency volatility using a stochastic volatility model with simultaneous and correlated jumps in returns and volatility. We estimate the model using an efficient sequential learning algorithm that allows for learning about multiple unknown model parameters simultaneously, with daily data on four popular cryptocurrencies. We find that these cryptocurrencies have quite different volatility dynamics. In particular, they exhibit different return-volatility relationships: While Ethereum and Litecoin show a negative relationship, Chainlink displays a positive one and interestingly, Bitcoin’s one changes from negative to positive in June 2016. We also provide evidence that the sequential learning algorithm helps better detect large jumps in the cryptocurrency market in real time. Overall, incorporating volatility jumps helps better capture the dynamic behavior of highly volatile cryptocurrencies.

Introduction

Cryptocurrencies are characterized by extreme tail risks—sudden, massive price swings that often defy standard financial modeling. Traditional econometric tools, such as GARCH or standard Stochastic Volatility (SV) models, frequently fail to capture the "jumps" inherent in these digital assets. Furthermore, most existing research relies on "batch" estimation methods like Markov Chain Monte Carlo (MCMC), which require re-processing the entire historical dataset every time a new day of trading occurs. This paper addresses these challenges by applying a sophisticated sequential learning algorithm to a model that accounts for simultaneous jumps in both prices and volatility, providing a real-time framework for understanding market risk.

Methodology: The SVCJ Model and Sequential Learning

The researchers employ the Stochastic Volatility model with Co-Jumps (SVCJ). Unlike simpler models that treat price jumps in isolation, the SVCJ framework assumes that jumps in returns and jumps in volatility are correlated and can occur simultaneously. This is critical for cryptocurrency markets, where a price crash often triggers an immediate and explosive increase in market variance.

To solve the estimation problem in a computationally efficient manner, the authors implement a sequential learning algorithm (specifically, a Particle Filter/Sequential Monte Carlo approach). This recursive method allows the model to "learn" and update its parameters—such as the leverage effect—as new data arrives daily. This represents a significant shift from static models, as it allows for the detection of structural changes in market behavior in real time without the need for full-sample re-estimation.

Data and Scope

The study utilizes daily closing prices (in USD) sourced from coinmarketcap.com for four major cryptocurrencies, representing a mix of established assets and newer tokens:

  • Bitcoin (BTC): April 2013 – August 2020
  • Litecoin (LTC): April 2013 – August 2020
  • Ethereum (ETH): August 2015 – August 2020
  • Chainlink (LINK): September 2017 – August 2020

Key Insights: The "Bitcoin Flip" and Market Heterogeneity

One of the paper's most significant findings is the evolution of the "leverage effect" in Bitcoin. In traditional equity markets, a negative correlation typically exists between returns and volatility: when prices fall, volatility rises. The authors found that Bitcoin followed this traditional pattern until June 2016. After this point, the relationship flipped; rising Bitcoin prices began to correlate with rising volatility. This suggests a transition in market regime, possibly driven by "noise" or uninformed traders entering the market during rallies.

Furthermore, the study reveals that the cryptocurrency market is not monolithic. While Ethereum and Litecoin maintained a traditional negative leverage effect, Chainlink exhibited a positive one throughout the sample period. These findings highlight that different tokens possess distinct volatility signatures, requiring tailored risk management strategies.

Empirical Results: Jump Intensity and Market Stress

The inclusion of volatility jumps proved essential for capturing the true risk profile of these assets:

  • Jump Frequency: Bitcoin’s jump intensity was measured at approximately 0.15 per trading day, equating to roughly 55 major jumps per year. This is significantly higher than the jump intensity of major US equity indexes, which typically experience only about 5 jumps per year (0.02 intensity).
  • The Pandemic Shock: The model’s real-time capabilities were demonstrated during the COVID-19 market crash on March 12, 2020. The algorithm captured a single "co-jump" that resulted in a 42% decline in Bitcoin’s daily return and a simultaneous 26.28% increase in return variance.
  • Annualized Volatility: The data confirms staggering risk levels across the board, with annualized volatility reaching 81.8% for BTC, 133.1% for ETH, and 150.8% for LINK.

Conclusion and Future Directions

The paper demonstrates that incorporating volatility jumps and utilizing sequential learning algorithms significantly improves the detection of market stress in real time. While the study uses finite-activity Poisson jumps, future research could explore infinite-activity Lévy jumps to capture more granular market noise. Additionally, investigating how simultaneous jumps across different cryptocurrencies impact the global leverage effect remains a promising area for further empirical study. Ultimately, the 2016 "Bitcoin flip" serves as a reminder that the underlying dynamics of digital assets are subject to rapid structural shifts as the investor base evolves.

Summary updated

Abstract

This paper studies the dynamics of cryptocurrency volatility using a stochastic volatility model with simultaneous and correlated jumps in returns and volatility. We estimate the model using an efficient sequential learning algorithm that allows for learning about multiple unknown model parameters simultaneously, with daily data on four popular cryptocurrencies. We find that these cryptocurrencies have quite different volatility dynamics. In particular, they exhibit different return-volatility relationships: While Ethereum and Litecoin show a negative relationship, Chainlink displays a positive one and interestingly, Bitcoin’s one changes from negative to positive in June 2016. We also provide evidence that the sequential learning algorithm helps better detect large jumps in the cryptocurrency market in real time. Overall, incorporating volatility jumps helps better capture the dynamic behavior of highly volatile cryptocurrencies.
BibTeX
@inproceedings{huang2021sequential,
  author = {Jing‐Zhi Huang and Zhijian Huang and Li Xu},
  title = {{Sequential Learning of Cryptocurrency Volatility Dynamics: Evidence Based on a Stochastic Volatility Model with Jumps in Returns and Volatility}},
  booktitle = {Quarterly},
  year = {2021},
  doi = {10.1142/s2010139221500105},
}
Figure