Speed-Limited Volatility Uncertainty: Time-Consistent Robust Pricing Without the Rectangularity Trap

When a trader does not know an asset's volatility and refuses to assume a probability model for it, the standard robust approach (Knightian uncertainty, meaning uncertainty about which probability model is even correct) prices claims by taking the worst case over all volatility paths that stay inside a band. In the most common version, Peng's $G$-Brownian motion, the adversary may jump volatility from the bottom of the band to the top at any instant and as often as it likes. This is not a whim of the modelers. It follows from pasting stability, the requirement that the set of candidate models stay closed when one switches from one model to another at any stopping time. That requirement is what makes today's robust price agree with tomorrow's recomputed price. The cost is well known in robust control as the "rectangularity trap": the worst case becomes frictionless adversarial switching that no market produces, and the resulting prices and hedges are very conservative.

We propose to place a speed limit on volatility. Candidate models are those in which volatility stays in the band and also changes no faster than a rate $L$. The key observation is that this set is not rectangular, yet it is still pasting-stable in the path-dependent sense of Nutz and van Handel. On continuous paths, the current instantaneous volatility can be read off the realized quadratic variation of the path. Pasting at a stopping time therefore only has to match the volatility already observed, and that one condition restores the dynamic programming principle. The parameter $L$ moves continuously between two existing objects. At $L=\infty$ we recover the $G$-expectation, which is time-consistent but over-conservative. At $L=0$ we recover the static band of constant-volatility Black–Scholes prices, which is tight but time-inconsistent. Every intermediate $L$ is time-consistent, prices with mutually singular models, and leads to a tractable equation on an enlarged state space.

Core Claim

Fix volatility bounds $\underline\sigma \le \bar\sigma$ and a speed limit $L \ge 0$. Let $\mathcal{P}^L$ be the set of laws of $dS_t = S_t\sigma_t\,dW_t$ for which every $\sigma_t \in [\underline\sigma,\bar\sigma]$ and $|\sigma_t-\sigma_s| \le L|t-s|$. Laws with different volatility paths are mutually singular, so $\mathcal{P}^L$ is nondominated. Define the local candidate set at time $t$ along path $\omega$ as all continuations whose volatility starts at the observed value $\sigma_t(\omega)$. With that definition, $\mathcal{P}^L$ satisfies the path-dependent concatenation condition of Constructing sublinear expectations on path space. The conditional sublinear expectation then obeys the tower property. Formally, the robust price of a European claim $g(S_T)$ is $V(0,S_0,\sigma_0)$, where $V$ solves

$$\partial_t V + \tfrac12\sigma^2 S^2\,\partial_{SS}V + L\,|\partial_\sigma V| = 0,\qquad V(T,S,\sigma)=g(S),$$

with reflection at $\sigma\in\{\underline\sigma,\bar\sigma\}$. If the current volatility is itself unknown within the band, the price takes an outer supremum over $\sigma_0$.

This replaces a fully nonlinear equation in the asset variable (the Black–Scholes–Barenblatt equation) with one that is linear in the second derivative in $S$ and nonlinear only in the first derivative in $\sigma$. The price paid is one extra state dimension. We claim three things:

  1. Speed-limited ambiguity is time-consistent and admits a superhedging duality.
  2. Its price band shrinks monotonically and continuously from the $G$-expectation band to the static band as $L$ falls.
  3. For realistic $L$ calibrated from realized-volatility dynamics, it achieves the same hedging coverage as the $G$-model with substantially less capital.

Relation to Prior Work

Three works motivate the proposal by documenting the over-conservatism of pasting stability. Superhedging and Dynamic Risk Measures under Volatility Uncertainty identifies pasting stability as the condition for the dynamic programming principle under volatility uncertainty and links the robust price to second-order backward stochastic differential equations. It works only with the frictionless rectangular set; we keep its superhedging interpretation but change the set. Time-Constrained Robust MDPs diagnoses the rectangularity trap in discrete robust control and imposes a Lipschitz drift constraint on physical parameters. That paper states that this breaks Bellman recursion and falls back on stacked-observation heuristics. Our claim is sharper and specific to finance: because volatility is observable from the path, the constrained set remains pasting-stable after state augmentation, so the dynamic programming principle and duality are exact. Policy Gradient Algorithms for Robust MDPs with Nonrectangular Uncertainty Sets attacks genuinely non-rectangular sets with Langevin and Frank–Wolfe methods whose cost grows with the uncertainty dimension. We avoid that cost by staying inside a time-consistent class.

Minimax theorems for American options in incomplete markets without time-consistency shows that pasting is not necessary for optimal-stopping minimax identities, but it requires a dominating reference measure. Our construction keeps pasting and abandons domination, so the two results are complementary rather than in conflict.

Several works supply methods. The path-dependent aggregation theory of Constructing sublinear expectations on path space gives the route to the tower property. The concatenation-stability characterizations of Coherent and convex monetary risk measures for bounded càdlàg processes suggest the right dual form. The Deep 2BSDE solver of Machine Learning Approximation Algorithms for High-Dimensional Fully Nonlinear PDEs and 2BSDEs, which parameterizes the solution's Hessian with neural networks and was demonstrated on a 100-dimensional Black–Scholes–Barenblatt equation, serves as both our numerical engine and our strongest baseline.

Research Program

Theory. The first theoretical task is to verify the measurability and analyticity conditions of Nutz–van Handel for $\mathcal{P}^L$. The delicate step is defining $\sigma_t$ as a path functional, a left limit of local realized variance, on a set of full measure under every element of $\mathcal{P}^L$. From there we derive the tower property, the superhedging duality with hedges in $S$ only, and a comparison principle showing that $V$ is monotone in $L$. We will also show that $V$ converges to the Black–Scholes–Barenblatt solution as $L\to\infty$ and to the static constant-volatility band as $L\to 0$.

Multi-asset extension. For several assets, the speed limit applies to a covariance-matrix path. The extra state dimension grows as $d(d+1)/2$, and we will study a low-rank factor parameterization to control this growth.

Computation. In one and two dimensions we solve the equation above on grids to obtain ground truth. In high dimensions the $\sigma$-direction carries no diffusion, so a backward-equation solver cannot recover $\partial_\sigma V$ from its martingale part. We will instead treat the adversary's volatility rate $u_t$, with $|u_t|\le L$, as a control, and train an adversary network jointly with a value or hedge network in a min–max Monte Carlo scheme. As a second route, we will add vanishing vol-of-vol noise and use a Deep 2BSDE-style solver on the augmented state.

Optimal stopping (secondary). We will combine the adversary with the stopping networks of Deep Optimal Stopping to price Bermudan claims under speed-limited ambiguity.

Baselines

Static constant-volatility band (naive reference). This prices each claim at the worst constant $\sigma\in[\underline\sigma,\bar\sigma]$ under Black–Scholes. It is the tightest cheap answer, but it is time-inconsistent: the worst constant volatility recomputed tomorrow can differ from today's, so hedges get abandoned midway. We will measure that inconsistency directly, as the gap between the price at time 0 and the expected recomputed price along simulated paths, and show that our method has zero gap by construction.

$G$-expectation price via Deep 2BSDE (strongest existing method). This is the Black–Scholes–Barenblatt price computed with Machine Learning Approximation Algorithms for High-Dimensional Fully Nonlinear PDEs and 2BSDEs. It is the strongest competitor because it is the only existing time-consistent, nondominated, high-dimensional method in the file. No reported numerical error was found on Lacuna for the 100-dimensional Black–Scholes–Barenblatt benchmark; the summary says only that the method "converged to a highly accurate solution within minutes." We will therefore rerun it ourselves on identical payoffs, dimensions, and volatility bands, with the same GPU budget and the same hyperparameter search size as our solver. Our method counts as beating it if two conditions hold. First, at matched superhedging coverage on held-out stochastic-volatility paths (for example, 99% of paths with no shortfall), it requires strictly less initial capital. Second, it converges to the $G$ price at large $L$, which confirms that the gain comes from the constraint and not from solver error.

Time-constrained adversarial training (TC-RMDP-style). This applies the stacked-observation adversarial approach of Time-Constrained Robust MDPs to hedging. Its Lipschitz-constrained adversary has no dynamic-programming guarantee. We will give it the same simulator, budget, and $L$. Our method beats it if it reaches equal or lower capital with smaller variance across seeds and zero time-inconsistency gap.

Ablation. To isolate the key ingredient, we will keep the $L$-constrained adversary but drop $\sigma$ from the state, so that time-$t$ prices are recomputed without conditioning on observed volatility. If this ablation reproduces our capital savings without inconsistency, the augmentation argument is unnecessary. If it shows inconsistency, the argument is doing real work.

Why not just…?

Why not just narrow the volatility band in the $G$-model? This is the most serious threat, because a narrower interval also lowers capital. It does so, however, by excluding volatility levels that do occur, whereas the speed limit keeps every level and excludes only impossible switching speeds. We will settle the question by tracing the capital-versus-coverage frontier on stochastic-volatility paths for both families: narrower bands under the $G$-model, and full band with varying $L$. If the narrowed-band frontier dominates ours, the proposal fails, and we will report that.

Why not just use one stochastic-volatility model such as Heston or SABR? A single model is dominated by one reference measure and assigns probabilities to volatility paths. It is exposed to misspecification of vol-of-vol, which is exactly the Knightian content. In our experiments, the simulator family includes models with parameters different from the one used to fit the hedge, and we report shortfall under that mismatch.

Why not just use model-free martingale optimal transport bounds from option smiles? Bounds for VIX Futures Given S&P 500 Smiles shows that smile information tightens bounds a great deal: on 2016 market data the classical lower bound of 0% rises to 8.4%, with an upper bound of 16.7%. Those bounds are static and say nothing about volatility dynamics. The two approaches are complementary, and we test this by intersecting the two: adding the speed limit to a smile-constrained problem in the synthetic SABR setting from that paper. We then check whether the reported 7.2%–22.8% band tightens further.

Proposed Experiments

Experiment 1: correctness in low dimensions. For European calls, digitals, and butterflies in one and two dimensions, the min–max solver must match grid solutions of the augmented equation across $L$. Its prices must lie between the static band and the $G$ band, and must be monotone in $L$.

Experiment 2: hedging under realistic volatility dynamics. We simulate Heston, SABR, and regime-switching volatility paths. We calibrate $L$ from a high quantile of their volatility rates, then compare capital, shortfall frequency, and expected shortfall across all baselines.

Experiment 3: scaling. We price basket options on the 100-dimensional Black–Scholes–Barenblatt benchmark configuration, using a low-rank volatility factor, and report wall-clock time against Deep 2BSDE under the same budget.

Experiment 4: Bermudan claims. We price Bermudan max-calls with the stopping networks of Pricing and Hedging American-Style Options with Deep Learning, and report how the $L$-band changes early-exercise value.

Decisive check. Two results would refute the central premise: the narrowed-band frontier dominating the speed-limit frontier in Experiment 2, or the no-state ablation matching our method with no measurable time-inconsistency. Either outcome would mean that state-augmented pasting stability buys nothing over simpler fixes.