The equivalent constant-elasticity-of-variance (CEV) volatility of the stochastic-alpha-beta-rho (SABR) model
Introduction
The stochastic-alpha-beta-rho (SABR) model proposed by Hagan et al. (2002) is one of the most popular stochastic volatility models adopted in the financial industry. Its commercial success is owing to a few factors. The model is intuitive and parsimonious. It provides a flexible choice of backbone, the trace of the at-the-money volatility against the spot price. Most importantly, Hagan et al. (2002) provide an analytic approximation of implied Black–Scholes (BS) volatility in closed form (hereafter, the HKLW formula), from which traders can readily convert to the option price using the BS formula.
The HKLW formula is an asymptotic expansion valid for small time to maturities (up to the first order in time) and near-the-money strike prices. Several authors have attempted to improve the HKLW formula. Based on the results of Berestycki et al. (2004), Obłój (2007) corrects the leading order term of the HKLW formula. Henry-Labordère (2005) derives the same leading order term from the heat kernel under hyperbolic geometry. Paulot (2015) further provides a second-order approximation which is accurate in a wider region of strike prices at the cost of numerical integrations for the second-order term. Further, Lorig et al. (2017) obtain implied BS volatility up to the third order in time, which unfortunately is valid only near the money. A more accurate solution of the SABR model, however, requires large-scale numerical methods such as the finite difference method (Hagan, Kumar, Lesniewski, Woodward, 2014, Park, 2014, von Sydow, Milovanović, Larsson, In’t Hout, Wiktorsson, Oosterlee, Shcherbakov, Wyns, Leitao, Jain, Haentjens, Waldén, 2019), continuous time Markov chain (Cui et al., 2018), multidimensional numerical integration (Antonov et al., 2013, Islah, 2009, Korn, Tang, 2013, Henry-Labordère), or Monte-Carlo simulation (Cai, Song, Chen, 2017, Chen, Oosterlee, Van Der Weide, 2012, Choi, Liu, Seo, 2019).
By nature, analytic approximation methods suffer two important drawbacks when some parameters or strike price go beyond the comfort zones of the concerned methods: non-negligible price error from the true value and the occurrence of arbitrage opportunity. Nevertheless, these methods are still attractive to practitioners because they are fast and robust. Note that practitioners need to compute the prices and Greeks of thousands of European options (or swaptions) frequently during trading hours. The calibration of the model parameters to the observed volatility smile also requires fast option evaluation because the parameters must be found using iterative methods. The numerical methods mentioned above are computationally intensive and not fast enough to use for those purposes.
Fortunately, the errors in analytic approximations are not a significant issue for those who use the SABR model primarily to price and manage the risk of European options. Specifically, the model parameters should first be calibrated to the market prices of the options at several liquid strike prices near the money. Then, the calibrated model is used to price the options at other strike prices. In this sense, the SABR model serves as a tool to interpolate (as well as extrapolate) the volatility smile, meaning that the accuracy of the price formula is less a concern.
Arbitrage under the analytic approximations occurs because an absorbing boundary condition is actually not imposed at the origin, as the small-time asymptotics of the transition density does not feel the boundary. The SABR process has a non-zero probability of hitting zero for , and an absorbing boundary condition should be explicitly imposed at the origin for for the price process to be a martingale and arbitrage-free. For this reason, the analytic approximations exhibit arbitrage opportunity in the low-strike region. The arbitrage outbreak is still an important concern to options market makers; savvy hedge funds can exploit them by purchasing a butterfly of options with a negative premium. To avoid such trades, market makers carefully keep track of the lower bound of the arbitrage-free region (and often patch a different arbitrage-free model below the bound). Therefore, the degree of arbitrage should be yet another performance measure for testing newly proposed analytic approximations, as in Obłój (2007), as much as the approximation error.
We contribute to the SABR model literature by proposing new analytic approximations that are more accurate and have a wider arbitrage-free strike region than existing studies. We derive the equivalent volatility under the constant-elasticity-of-variance (CEV) model, from which the option price is computed with the analytic CEV option price formula (Schroder, 1989). We provide two formulas for the equivalent CEV volatility as spin-offs from existing studies. The first one (Theorem 1) is obtained by following the approximation method of Hagan et al. (2002). The second one (Theorem 2) is simplified from Paulot (2015)’s original CEV volatility formula.
Our CEV-based approach is motivated by the simple intuition that the SABR model should converge to the CEV model when the volatility of volatility (vol-of-vol) approaches zero. Such motivation for using the CEV model is not novel in the SABR model literature. Yang et al. (2017) show that the CEV option price (with the CEV volatility being the initial SABR volatility) is a good approximation in certain parameter ranges and is, naturally, arbitrage-free. The practical use of the result, however, is limited because the parameters related to the volatility process (i.e., vol-of-vol and correlation) are ignored in the approximation and only one degree of freedom is left to fit the volatility smile. Our work extends Yang et al. (2017) as our CEV approximations have full dependency on the SABR parameters. Paulot (2015, §3.6, 4.5) also discusses the equivalent CEV volatility as an alternative to BS volatility and outlines the derivation. His emphasis, however, is placed on BS volatility and the CEV volatility approximation is not tested numerically. Further, there is no discussion about the implications such as the mass at zero. In short, the advantage of the CEV approach has not thus far been explored. We fill this research gap by advocating the use of CEV volatility for the SABR model.
The numerical results show that our CEV-based approximations are more accurate than the corresponding BS-based methods from which they stem. In particular, the presented CEV approaches are more accurate when the initial volatility is large. This finding complements Paulot (2015)’s refinement, which makes the approximation more accurate for large vol-of-vol. Having both advantages, the second CEV approximation based on Paulot (2015) performs the best among all the methods over wide parameter ranges. In the numerical test for comparing the degree of arbitrage, the second CEV approximation also performs the best among all the methods in that negative implied probability density starts to appear at the lowest strike price (see Section 4.3).
Surprisingly, the projection of the SABR model to the CEV model has the effect of imposing an absorbing boundary condition at the origin because the CEV price formula assumes the same boundary condition. Our CEV approximations offer finite CEV volatility at a zero strike, making them capable of implying the probability of hitting the origin. The mass-at-zero approximation in a closed-form formula (Theorem 3) shows excellent agreement with the numerical results in small time and is consistent with the exponentially vanishing asymptotics in the limit (Chen and Yang, 2019). To the best of our knowledge, our approximation is the first closed-form approximation that works for all , as the existing estimation methods either work on the uncorrelated case only (Gulisashvili et al., 2018) or require numerical integration for the correlated case (Yang and Wan, 2018). Even if the mass at zero from our CEV approximations becomes less accurate in large time or vov-of-vol, our CEV approximations remain internally consistent with the model-free smile shape determined by the (possibly incorrect) mass at zero (De Marco et al., 2017). This explains why our CEV approach has a wider arbitrage-free region.
The remainder of this paper is organized as follows. Section 2 reviews the SABR model and existing BS volatility approximations. Section 3 derives the equivalent CEV volatility and mass-at-zero approximation. Section 4 presents the numerical results. Finally, Section 5 concludes.
Section snippets
SABR model and analytic approximations
In this section, we introduce the SABR model and review various analytic approximation methods for the equivalent BS volatility in the order of increasing accuracy. Rather than simply repeating existing results, we reorganize the formulas in an insightful way, which should lead to our new approach in Section 3.
CEV model
Since we advocate the use of implied CEV volatility, we briefly review the CEV model. The standardized CEV model with volatility is given byThe standardized prices of the call and put options with strike price and time-to-maturity are respectively (Schroder, 1989)where and are
Numerical results
In this section, we numerically test the two CEV-based approximations in comparison to existing BS-based approximations.5 For easier reference, the methods are labeled as follows:
- •
BS-A: Eq. (9), the HKLW formula of Hagan et al. (2002).
- •
BS-B: Eq. (23), Paulot (2015)’s equivalent BS volatility of order .
- •
BS-C: Lorig et al. (2017, §5.4)’s equivalent BS volatility up to
- •
DMHJ: Eq. (24), the
Conclusion
The SABR model is the dominant stochastic volatility model in the financial industry. Since finding an accurate solution is computationally burdensome, robust analytic approximations of the equivalent volatility are favored in practice despite their imperfection. We show that the quality of the approximation can be significantly improved by deriving the equivalent CEV volatility instead of the BS volatility. Projecting the SABR model on the CEV model takes advantage of an absorbing boundary
Acknowledgements
Jaehyuk Choi was supported by the 2019 Bridge Trust Asset Management Research Fund. Lixin Wu was supported by Grant #16306717 of the Research Grants Council of Hong Kong. The authors thank the editor and two anonymous referees for their helpful comments. We also thank Nan Chen and Nian Yang for providing the data points of Chen and Yang (2019, Figure 1).
References (37)
- et al.
The survival probability of the SABR model: asymptotics and application
Quantitative Finance
(2018) - et al.
SABR spreads its wings
Risk
(2013) - et al.
The free boundary SABR: natural extension to negative rates
Risk
(2015) - et al.
Modern SABR analytics
(2019) Théorie de la spéculation
Annales Scientifiques de l’École Normale Supérieure
(1900)- et al.
Computing the implied volatility in stochastic volatility models
Commun Pure Appl Math
(2004) - et al.
Exact simulation of the SABR model
Oper Res
(2017) - et al.
A low-bias simulation scheme for the SABR stochastic volatility model
International Journal of Theoretical and Applied Finance
(2012) - et al.
The principle of not feeling the boundary for the SABR model
Quantitative Finance
(2019)
Hyperbolic normal stochastic volatility model
Journal of Futures Markets
A note on the option price and ‘mass at zero in the uncorrelated SABR model and implied volatility asymptotics’
Quantitative Finance (forthcoming)
A general valuation framework for SABR and stochastic local volatility models
SIAM Journal on Financial Mathematics
Full-fledged SABR through markov chains
Wilmott
Shapes of implied volatility with positive mass at zero
SIAM Journal on Financial Mathematics
An approximation for the noncentral chi-squared distribution
Communications in Statistics - Simulation and Computation
Mass at zero in the uncorrelated SABR model and implied volatility asymptotics
Quantitative Finance
Arbitrage-free SABR
Wilmott
Cited by (4)
A Black–Scholes user's guide to the Bachelier model
2022, Journal of Futures Markets