The SABR model is a standard two-equation description of how an asset price and its volatility move together, and traders use it widely to fit option smiles. Pirjol and Zhu proved that in a common parameter setting the model has a hidden defect. When the asset falls, volatility rises, and the price's sensitivity to volatility is a power between 0 and 1. In that setting the instantaneous volatility can blow up to infinity in finite time with positive probability. The model then assigns infinite prices to VIX futures and VIX calls, the VIX being the market's 30-day volatility index (VIX options in the SABR model). Their fix is to cap the volatility process between two bounds $a$ and $b$. The fix has two weaknesses. The bounds come with no rule for choosing them, and the short-maturity formula they derive gets the curvature of the VIX smile wrong. Nobody has checked any of this against market quotes.
We propose to run that check, borrowing a discipline from engineering validation: when a simulation disagrees with an experiment, identify a physical correction, measure it independently, and confirm that it closes the gap out of sample. We treat the volatility cap as such a correction. We infer it from one market instrument, the VIX futures level, and test whether it predicts another, the VIX option smile across maturities. We price everything exactly rather than through the short-maturity formula. The result answers two questions. Is the reported curvature mismatch a real failure of capped SABR, or an artifact of the asymptotic formula used to compute it? And is the cap a stable, identifiable quantity, or an arbitrary knob?
Core Claim
Our hypothesis is that the cap $b$ is an economically meaningful parameter. It sets how much of the would-be explosion the market actually prices, and three things should follow. First, $b$ should be pinned down by VIX futures, because under explosive parameters the futures price is dominated by paths that reach the cap. Second, $b$ calibrated on futures should predict out-of-the-money VIX call prices without refitting. Third, the curvature mismatch should shrink substantially when capped SABR is priced exactly at observed maturities of 7 to 126 days, rather than through the zero-maturity expansion. If any of these fails, we learn something sharper than "capped SABR is imperfect": we learn which failure it is. A $b$ that drifts erratically day to day says the cap is a nuisance parameter. A curvature mismatch that survives exact pricing says the capped dynamics are structurally wrong for VIX. A simpler model that avoids explosion entirely while fitting as well says the cap was never needed.
The study's novelty is a falsifiable test, not a new model. We test whether a mathematical pathology proved in the theory literature has an observable footprint in prices, and whether its proposed remedy survives contact with data.
Research Program
The model writes the forward price as $dX_t = \alpha_t X_t^\beta dW_{1,t}$ and its volatility level as $d\alpha_t = \nu \alpha_t dW_{2,t}$. The two noises have correlation $\rho$. The instantaneous return volatility is $v_t = \alpha_t X_t^{\beta-1}$. When $\rho<0$ and $0<\beta<1$, a falling $X_t$ raises $\alpha_t$ and also raises $X_t^{\beta-1}$, and this compounding feedback is what Feller's test, a classical criterion for whether a one-dimensional diffusion reaches infinity, detects as explosion (VIX options in the SABR model). The same pathology appears in interest-rate models with power-law volatility. There, Pirjol shows explosion with positive or full probability using Lyapunov and Khasminskii arguments (Explosion in the quasi-Gaussian HJM model). He also shows that futures prices blow up strictly before the underlying state does (Small-noise limit of the quasi-Gaussian log-normal HJM model). That last result motivates our focus on VIX futures as the instrument most sensitive to the cap.
The study has four steps.
Step 1: an exact pricer. The model is Markovian in the pair $(X,\alpha)$, so the squared VIX at a future date is a function of those two numbers: the expected integrated variance over the next 30 days. We compute that function once by solving the backward pricing equation on a two-dimensional grid. VIX futures and options at each maturity then follow from the joint distribution of $(X_T,\alpha_T)$, obtained by a second grid solve or by Monte Carlo with that function plugged in.
The grid near $X=0$ is the delicate part. There we use the weighted finite-element construction of Dirichlet Forms and Finite Element Methods for the SABR Model, which supplies a convergent scheme for SABR's degenerate boundary. Its convergence proof assumes $|p|\nu^2<2$, so we report separately any calibrations that land outside that range. For the equity side we check the joint density against the closed-form expansion of SERIES EXPANSION OF THE SABR JOINT DENSITY, which reports pointwise errors of 0.2% to 2% against finite differences. Its small vol-of-vol assumption means we expect agreement only at short maturities.
Step 2: where does the market sit? We calibrate uncapped SABR daily to SPX option smiles and record how often the fitted parameters fall in the explosive region. For those days we estimate how likely volatility is to exceed very large levels within 30 days. This step addresses a gap neither line of work has touched: whether real fitted parameters actually live where the theory says the model breaks.
Step 3: identify and validate the cap. We fix the SPX-calibrated parameters and calibrate only $b$ to the VIX futures term structure. The lower bound $a$ is set by the historical floor of the VIX and tested for sensitivity. We then predict the VIX option smiles out of sample.
As a contrast, we also fit all parameters jointly to SPX plus VIX. For this six-dimensional search we use Bayesian optimization with a tree-structured Parzen estimator, following Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization. That work showed simultaneous global search resolves coupled parameter dependencies that sequential fitting misses. We compare it against multi-start gradient fitting with the same number of pricer calls.
Step 4: decompose the curvature error. For every maturity we compute market curvature, exact capped-model curvature, and the asymptotic formula's curvature. This splits the total mismatch into the part due to the formula and the part due to the model. The decomposition plays the same role as the bellows study (Mechanism of Mechanical Analysis on Torsional Buckling of U-Shaped Bellows). There, a 3.3% torque error fell to 0.9% once a measured wall-thinning defect was included. The lesson was that the residual error had a nameable physical source, and we look for the analogous source here.
Baselines
The naive reference is uncapped SABR priced by Euler Monte Carlo at a typical practitioner step size. It always returns finite VIX prices, because a discrete simulation cannot reach infinity. We will show those prices drift upward without converging as the step shrinks. That would expose the discretization as an undeclared cap, and an uncontrolled one. Any reported fit from this baseline is therefore an artifact. It still serves as the floor for what "working" SABR looks like in practice.
The strongest existing method is the modulated rough Volterra model of Volatility Options in Rough Volatility Models. It adds an independent random factor to a rough forward-variance model and was calibrated to CBOE VIX options on May 14, 2014, at 7, 35, 63, 98, and 126 days. It reports pricing errors "below 3 cents (USD) for VIX options." It is the only model in the file with a reported fit to real VIX quotes, so it is the benchmark.
For a fair comparison we use the same date and maturities and quote errors in the same dollar units. We give capped SABR no more free parameters per maturity than the Volterra model uses. We report single-maturity and joint-maturity fits separately, since the Volterra authors note their joint fits degrade. Beating it means one of two things. The first is staying below 3 cents per maturity with fewer parameters. The second is a lower joint-maturity error with parameters that stay stable across days. Matching it at a fraction of the computational cost counts as a partial win, but not beating it.
The structurally simplest non-explosive competitor is SABR with $\beta=1$, where the feedback mechanism disappears. It is fit with the same pricer and budget.
Two ablations isolate the key ingredient. The first frees $b$ per maturity instead of fixing it from futures; the gap between the two measures how much predictive content the cap really has. The second replaces the hard cap with a smooth damping of the volatility coefficient. That tests whether the curvature problem comes from the kink at the boundary.
For the SPX side, Deep Smoothing of the Implied Volatility Surface is the surface-fit reference on the 2008 and 2018 windows, a hybrid of a parametric surface and a neural corrector with soft no-arbitrage penalties. No reported result for its RMSE or MAPE values was found on Lacuna; they appear only graphically. We will rerun it on our dates as an upper bound on how well any surface can fit, not as a dynamic competitor.
Why not just…?
Why not just set β = 1 and avoid explosion entirely? This is the most serious threat. If lognormal SABR fits SPX and VIX jointly as well as capped SABR with $0<\beta<1$, the cap is unnecessary and the honest conclusion is "don't use the explosive region." We include it as a full baseline with an identical budget. We expect it to struggle to produce both the SPX skew's shape and the VIX futures level, because $\beta<1$ is usually needed for the former. That is an expectation, not evidence, and the experiment decides it.
Why not just use the rough Volterra model, which already fits to under 3 cents? On a single date at separate maturities it may simply win, and we will report that. Its acknowledged weaknesses are joint-maturity instability and costly non-Markovian simulation. Capped SABR's two-state structure allows grid pricing, which is cheap enough for daily stability studies. The comparison that matters is the joint-maturity fit and parameter stability over many dates, which the Volterra paper does not report.
Why not just restrict to short maturities where explosion is unlikely? Futures blow up before the state variable does (Small-noise limit of the quasi-Gaussian log-normal HJM model). Moreover, the VIX at any date already integrates over a 30-day window, so even the shortest VIX contract is exposed. Step 2 measures how much explosion probability actually accumulates within 30 days under fitted parameters. If that probability is negligible, this alternative is vindicated, and we will say so.
Why not just fit a neural surface or a GARCH model? A neural surface interpolates prices but recovers no dynamics, so it cannot say anything about explosion (Deep Smoothing of the Implied Volatility Surface). GARCH-type models and their neural successors, such as A Neural Stochastic Volatility Model, describe historical returns rather than pricing probabilities. They answer a different question.
Proposed Experiments
The data are CBOE VIX futures and options quotes for May 14, 2014, matching the Volterra benchmark, plus daily quotes in the 2008 crisis window and the 2018 calm window used by the deep-smoothing study. We pair these with SPX option smiles on the same days. Licensed CBOE data are required, and the file lists no public copy.
Five checks are decisive. The first is the Euler step-refinement test on uncapped SABR. Non-convergence confirms that practitioner simulations hide an implicit cap. The second is the fraction of daily SPX calibrations in the explosive region and their 30-day explosion probability. The third is out-of-sample VIX option error when $b$ is fit only to futures, compared with the 3-cent benchmark and with $\beta=1$. The fourth is the curvature decomposition at each maturity. A model-attributable share near zero means the published mismatch was a formula artifact, and a large share means capped SABR is structurally wrong for VIX. The fifth is the day-to-day stability of the fitted $b$, compared between 2008 and 2018. A cap that tracks a stable function of market stress supports the claim that it is a real quantity. A cap that jumps randomly refutes it.
Each outcome is informative. Even the negative ones say whether practitioners should abandon the explosive region of SABR for volatility products or keep it with a validated remedy.
References
- VIX options in the SABR model
- Explosion in the quasi-Gaussian HJM model
- Small-noise limit of the quasi-Gaussian log-normal HJM model
- Dirichlet Forms and Finite Element Methods for the SABR Model
- SERIES EXPANSION OF THE SABR JOINT DENSITY
- Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization: A Case Study on Pebax Polymer
- Mechanism of Mechanical Analysis on Torsional Buckling of U-Shaped Bellows in FLNG Cryogenic Hoses
- Volatility Options in Rough Volatility Models
- Deep Smoothing of the Implied Volatility Surface
- A Neural Stochastic Volatility Model