Heston vol-of-vol and the VVIX
Jherek Healy · 2025
Read Heston vol-of-vol and the VVIX by Jherek Healy (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-19611.txt.
Introduction
Finance researcher Jherek Healy examines a calibration headache inside the Heston model, the most widely used stochastic volatility model for pricing and risk-managing exotic derivatives. Even a popular model can be hard to fit to market prices, and unstable fits make exotic option prices unreliable. Healy zeroes in on one parameter, the vol-of-vol, and asks how it connects to the VVIX, the market index of volatility-of-volatility.
Four ways of estimating the VVIX inside the Heston model are laid out side by side. Two rest on the known transition density of the variance process, one is an analytical approximation, and one solves the Heston partial differential equation using the underlying S&P 500 directly. Healy then tests whether these estimates can steady the calibration process.
Catalog record
| CREATOR | Jherek Healy |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 25 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-19611
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Heston vol-of-vol and the VVIX Jherek Healy Abstract The Heston stochastic volatility model is arguably, the most popular stochastic volatility model used to price and risk manage exotic derivatives. In spite of this, it is not necessarily easy to calibrate to the market and obtain stable exotic option prices with this model. This paper focuses on the vol-of-vol parameter and its relation with the volatility of volatility index (VVIX) level. Four different approaches to estimate the VVIX in the Heston model are presented: two based on the known transition density of the variance, one analytical approximation, and one based on the Heston PDE which computes the value directly out of the underlying SPX500. Finally we explore their use to improve calibration stability. keywords stochastic volatility; Heston; vol-of-vol; VVIX; PDE † † firstpage: 1 † † issue: 1 † † articlenumber: 1 † † year: 2024 † † copyright-year: 2024 † † authornames: Jherek Healy † † address: [email protected] \history 1 Introduction Under the Heston model 8 , the asset X follows \displaystyle\mathop{}\!\kern 0.0pt\mathrm{d}X(t) \displaystyle=(r(t)-q(t))X(t)dt+\sqrt{V(t)}X(t)\mathop{}\!\kern 0.0pt\mathrm{d}W_{X}(t)\,, (1a) \displaystyle\mathop{}\!\kern 0.0pt\mathrm{d}V(t) \displaystyle=\kappa\left(\theta-V(t)\right)+\sigma\sqrt{V(t)}\mathop{}\!\kern 0.0pt\mathrm{d}W_{V}(t)\,, (1b) with W_{X} and W_{V} being two Brownian motions with correlation \rho , and r,q the instantaneous growth and dividend rates. The model, due to its affine properties, leads to an explicit characteristic function, enabling relatively fast pricing of European options. As a consequence, it has been extremely popular in the literature, as well as for real world use in banks, in order to risk manage the price of exotic derivatives, such as Cliquet options 5 ; 7 ; 4 . While it is relatively straightforward to calibrate the model against a set of vanilla options, it is difficult to obtain a stable, meaningful calibration. Indeed, a change of objective function used in the optimization, or the use of different tactics for calibration, such as reducing the number of parameters lead to vastly different prices for exotics. This is well described in 6 , where the two of the model parameters are set exogenously, v(0) is set to the VIX value and \theta is set from a moving window estimate of the historical VIX quotes. Some practitioners also fix the speed of mean reversion \kappa exogenously. Yet another practice is to forget the consistency across deals, and calibrate only against a single maturity T , setting v_{0}=\theta and a specific term-structure of \kappa ( 3 recommends 5/T ). The VVIX is an index measuring the volatility of the VIX index, which is itself a volatility index. The VVIX thus measures the vol of vol. We thus may use the VVIX quote as vol-of-vol of the Heston model. But does the vol of vol \sigma of the Heston model really corresponds to the VVIX? Does it help stabilizing the calibration in some sense? This paper intends to answer those questions, by analyzing the theoretical VVIX price in the Heston model. 2 Approximate VVIX in the Heston model The VIX index corresponds to the square root of a variance swap replication for a maturity of 30 days, using a simple discretization scheme, based on liquid vanilla option prices on the SPX500 index: \mathrm{VIX}(t,t+\Delta)^{2}\Delta=-\left(1-\frac{F(t,t+\Delta)}{K^{\star}}\right)^{2}+2\sum_{i=1}^{n}w_{i}V_{\textmd{OTM}}(K_{i},t,t+\Delta)\,, (2) where w_{i}=\frac{K_{i+1}-K_{i}}{2K_{i}^{2}} for i=2,...,n-1 , w_{1}=\frac{K_{2}-K_{1}}{2K_{1}^{2}} , w_{n}=\frac{K_{n}-K_{n-1}}{2K_{n}^{2}} , \Delta is 30 days (we will use the ACT/365 convention), V_{\textmd{OTM}}(K_{i},t,T) is the market price of an out-of-the-money option of strike K_{i} and maturity T . K^{\star} is the first strike K_{i} below F(t,t+\Delta) . We will approximate the VIX square by the true value of the expected 30-day variance in the Heston model \mathrm{VIX}(t,T)^{2}\approx\frac{1}{T-t}\operatorname{E}\left[\int_{t}^{T}v(u)\mathop{}\!\kern 0.0pt\mathrm{d}u\nonscript\;\middle|\nonscript\;t\right]\,. (3) In the Heston model, the above expectation is known explicitly and reads \mathrm{VIX}_{H}(v(t),T-t)^{2}=\frac{1}{T-t}\operatorname{E}\left[\int_{t}^{T}v(u)\mathop{}\!\kern 0.0pt\mathrm{d}u\nonscript\;\middle|\nonscript\;t\right]=\left(1-\frac{1-e^{-\kappa(T-t)}}{\kappa(T-t)}\right)\theta+\frac{1-e^{-\kappa(T-t)}}{\kappa(T-t)}v(t)\,. (4) The buyer of a VIX future of maturity T will receive at maturity, the VIX index value computed from option prices of maturity T+\Delta . Using the expected variance as VIX square (Equation 3 ) leads to F_{\mathrm{VIX}}(t,T)\approx\operatorname{E}\left[\sqrt{\mathrm{VIX}(T,T+\Delta)^{2}}\nonscript\;\middle|\nonscript\;t\right]\,. (5) Similarly, the VVIX corresponds to a variance swap replication for a maturity of 30 days, based on the VIX index. Options on the VIX index of maturity T , are actually options on a VIX future with maturity T , and thus measuring the volatility from T to T + 30 days, the 30 day forward volatility. It may be approximated by the log contract value 1 : \mathrm{VVIX}(t,T)^{2}\approx\frac{2}{T-t}\operatorname{E}\left[\ln{F_{\mathrm{VIX}}(T,T)}-\ln F_{\mathrm{VIX}}(t,T)\nonscript\;\middle|\nonscript\;t\right] (6) For a CIR process, the transition law is known explicitly 2 and reads v(t)=\frac{\sigma^{2}\left(1-e^{-\kappa(t-u)}\right)}{4\kappa}\chi_{d}^{\prime 2}\left(\frac{4\kappa e^{-\kappa(t-u)}}{\sigma^{2}\left(1-e^{-\kappa(t-u)}\right)}v(u)\right)\,, (7) where \chi_{d}^{\prime 2} denotes the non-central chi-squared distribution with d degrees of freedom where d=\frac{4\kappa\theta}{\sigma^{2}} . Let C_{1}=\frac{\sigma^{2}\left(1-e^{-\kappa\Delta}\right)}{4\kappa} , \phi_{t,u}(z) be the probability density function of the non-central chi-squared distribution, combining Equation 4 with Equation 6 , we may thus compute the VVIX using a one-dimensional integration. \mathrm{VVIX}(t,T)^{2}\approx\frac{2}{T-t}\int_{0}^{\infty}(\ln\sqrt{\mathrm{VIX}_{H}(v(T)=C_{1}z,\Delta)^{2}}-\ln F_{\mathrm{VIX}}(t,T))\phi_{t,T}(z)\mathop{}\!\kern 0.0pt\mathrm{d}z (8) with F_{\mathrm{VIX}}(t,T)=\int_{0}^{\infty}\sqrt{\mathrm{VIX}_{H}(v(T)=C_{1}z,\Delta)^{2}}\phi_{t,T}(z)\mathop{}\!\kern 0.0pt\mathrm{d}z (9) where the expectation inside each integral is given explicitly by Equation 4 . The VIX future formula corresponds to the one given in 15 . When the Feller condition \kappa\theta\geq\sigma^{2} does not hold, the integral is more challenging to calculate due to the explosion at zero. We found only minor discrepancies using Julia’s QuadGK 1 1 1 https://github.com/JuliaMath/QuadGK.jl . Alternatively, 16 present a Fourier based approach to compute the expectation, using the characteristic function of the variance process, which may be more stable numerically and better performing. In practice, the VVIX is calculated using only a discrete set of liquid options on the VIX: \mathrm{VVIX}(t,T)^{2}(T-t)=-\left(1-\frac{F_{\mathrm{VIX}}(t,T)}{K^{V\star}}\right)^{2}+\sum_{i=1}^{n}w_{i}V_{OTM}(K^{V}_{i},t,T) (10) where w_{i}=\frac{K^{V}_{i+1}-K^{V}_{i}}{2{K^{V}_{i}}^{2}} for i=2,...,n-1 , w_{1}=\frac{K^{V}_{2}-K^{V}_{1}}{2{K^{V}_{1}}^{2}} , w_{n}=\frac{K^{V}_{n}-K^{V}_{n-1}}{2{K^{V}_{n}}^{2}} , K^{V\star} is the first strike K^{V}_{i} below F_{VIX} . The price of VIX option can be computed using a one-dimensional integration, very much like for the continous VVIX approximation: V(K,t,T)=\int_{0}^{\infty}\max\left(\eta\sqrt{\mathrm{VIX}_{H}(v(T)=C_{1}z,\Delta)^{2}}-\eta K,0\right)\phi_{t,T}(z)\mathop{}\!\kern 0.0pt\mathrm{d}z (11) with \eta=1 for a call and \eta=-1 for a put option. A practical implementation will truncate the integration at the level z where the intrinsic value is zero in order to keep the integrand smooth. If we assume that the VIX option prices are given by Equation 11 , the continuous approximation (Equation 8 ) is remarkably close to the more exact VVIX replication, as long as the lowest replication strike is small enough. A typical strike range is [10,65] for a VIX=14. In particular, the lowest strike is not that small compared to the VIX value. This creates a bias in the VVIX value, which may be particularly pronounced if the VVIX is large (Table 1 with the first VIX option strike K^{V}_{1}=5 or K^{V}_{1}=10 ). Table 1: Effect of the VIX option strike range truncation on the estimated VVIX value, compared to the log contract price, for different Heston parameters. Set Heston parameters F_{\mathrm{VIX}} Log Replication Simple v_{0} \kappa \theta \rho \sigma contract K^{V}_{1}=5 K^{V}_{1}=10 I 0.0236 0.2575 0.0849 -0.7513 0.3150 15.4 105.4 105.2 (0.2%) 100.0 (5.1%) 98 II 0.0313 0.75 0.0678 -0.7663 0.7593 16.0 218.5 218.4 (0.0%) 193.0 (11.5%) 204 III 0.0371 3.4490 0.0497 -0.7558 1.7522 15.9 231.8 229.8 (0.9%) 225.9 (2.5%) 382 IV 0.0538 0.6431 0.0880 -0.7010 0.6159 22.2 143.7 143.8 139.8 127 V 0.0440 0.75 0.0998 -0.7410 0.7654 19.9 184.8 184.8 176.8 171 VI 0.0397 4.6705 0.0696 -0.7149 2.0640 18.9 196.7 196.5 196.7 376 We also derive a simple approximation in Appendix A , which does not require any numerical integral calculation, but is not particularly accurate, especially for large speeds of