Abstract
This paper investigates the continuous-time dynamics of the VIX index by incorporating stochastic volatility and simultaneous jumps in both the VIX level and its volatility factor. Building upon a general parametric affine model, we derive a linear relationship between the latent stochastic volatility factor and the market-observed VVIX index. This theoretical link allows the VVIX to serve as a direct proxy for volatility, simplifying the estimation process. Using Markov Chain Monte Carlo (MCMC) methods on data spanning 2007 to 2014, we detect significant co-jumps between the VIX and VVIX. Our findings demonstrate that a "double-jump" stochastic volatility model (SVJJ) with state-dependent jump intensity significantly outperforms nested models, providing a more accurate representation of market stress and the "volatility of volatility."
Introduction
The VIX index, widely recognized as the market's "fear gauge," measures the 30-day expected volatility of the S&P 500. While the VIX is itself a measure of volatility, it exhibits its own significant fluctuations—often referred to as the "volatility of volatility." Accurately modeling these dynamics is essential for the pricing and risk management of VIX-linked derivatives, such as futures and options.
Historically, modeling the VIX has faced two primary hurdles. First, the volatility of the VIX is a latent (unobservable) variable, making it difficult to estimate accurately using only VIX price data. Second, the computational burden of pricing VIX options within complex stochastic volatility frameworks often requires intensive numerical methods. This research addresses these challenges by utilizing the VVIX index—a measure of the VIX's implied volatility—to ground the estimation of a new double-jump stochastic volatility model.
The SVJJ Model and the VVIX Proxy
The authors propose a Double-Jump Stochastic Volatility (SVJJ) model. This framework assumes that the logarithm of the VIX follows an affine jump-diffusion process. The defining innovation of this paper is the derivation of a near-linear relationship between the squared VVIX index and the latent stochastic volatility factor ($\omega_t$).
By establishing this link, the VVIX index transforms from a mere market indicator into a powerful proxy for the model's hidden volatility component. This allows the researchers to ensure that their estimated volatility remains consistent with market-implied expectations, bridging the gap between physical price dynamics and option-market sentiment.
Methodology
To estimate the model parameters and identify jump characteristics, the study employs a multi-step empirical approach:
- Co-jump Detection: Using a non-parametric test, the authors identify instances where the VIX and its volatility jump simultaneously. This helps justify the inclusion of a "double-jump" component in the model.
- MCMC Estimation: The researchers use Markov Chain Monte Carlo (MCMC) methods to estimate the model. This Bayesian approach is particularly effective for models with latent variables and jump processes, as it provides posterior distributions for jump times and magnitudes rather than simple point estimates.
- Comparative Analysis: The study compares the proposed SVJJ model against simpler nested models, such as standard stochastic volatility (SV) and models with jumps only in the VIX level (SVJ).
Data and Empirical Evidence
The model was tested using daily closing data for the VIX and VVIX indices from January 3, 2007, to November 26, 2014. This period is particularly significant as it encompasses the calm pre-crisis era, the extreme volatility of the 2008 financial crisis, and the subsequent market recovery.
Out of the 1,939-day sample, the authors detected 131 days of significant co-jumps between the VIX and VVIX. These events typically coincide with major macroeconomic shocks, where both the level of fear (VIX) and the uncertainty regarding that fear (VVIX) spike in tandem.
Key Insights
1. The Persistence of Volatility Jumps
Unlike standard diffusion-based shocks, which tend to mean-revert quickly, jumps in volatility shift the VIX into a new regime. For instance, during the onset of a crisis, a jump in the volatility factor causes the VIX to remain elevated and highly active for an extended period. Models that ignore these volatility jumps fail to capture this "regime-shifting" behavior.
2. State-Dependent Jump Intensity
The research finds that jumps are not purely random; they are state-dependent. When market volatility is already high, the probability of experiencing further jumps increases. This "jump clustering" suggests that market stress is self-reinforcing during periods of high turbulence.
3. Superior Model Fit
The SVJJ-S model (Double-jump with stochastic intensity) provided the best statistical fit. By accounting for jumps in both the VIX and its volatility, the model significantly reduced the "fat tails" in the residual error distribution. The correlation ($\rho$) between VIX shocks and volatility shocks was estimated at approximately 0.52, confirming a strong positive relationship: when the VIX rises, its volatility typically rises as well.
Conclusion
This paper provides strong evidence that jumps in both the VIX and its underlying volatility factor are essential for capturing the index's empirical behavior. By leveraging the VVIX index as a proxy, the authors offer a computationally tractable way to align stochastic volatility models with market reality. These findings have significant implications for traders and risk managers, suggesting that failure to account for "double jumps" and state-dependent intensity can lead to a substantial underestimation of risk during market crises.
Future research may further refine these models by exploring the specific risk premia that connect these physical dynamics to the pricing of VIX options and futures.
Summary updated