Drift-Aware Radii for Robust Superhedging

A robust superhedging price is the cheapest amount of starting capital that, with the right trading strategy, covers an option's payoff under every plausible model of the market. Recent work makes "plausible" concrete. It takes the empirical distribution of prices, draws a ball of fixed radius around it in Wasserstein distance (a measure of how much probability mass must be moved to turn one distribution into another), and computes the worst case over price processes inside that ball. The radius is set to cover sampling error alone. Markets drift, however, and a window of historical data describes a distribution that is already out of date. A radius sized only for sampling noise therefore under-hedges when drift is fast. A radius widened by a fixed safety factor over-hedges when drift is slow and returns bounds too wide to use.

We propose to size the radius from two measured quantities: the sampling error of the window and the observed speed of drift. Balancing the two gives a closed-form rule for both window length and radius, $\varepsilon^\star \propto \delta^{1/3}$, where $\delta$ is the per-period drift rate. The same rule predicts a sharp threshold. Below a critical drift rate the robust price stays strictly tighter than classical model-free bounds. Above it, the price collapses to those bounds and carries no information. This is the financial counterpart of the "small data" condition for nonlinear PDEs, under which good behaviour is guaranteed only while the perturbation stays below a quantified threshold. The rule needs no labels or realized hedge outcomes. It uses only the stream of observed prices, so it sidesteps the long feedback delay that makes learning-based tuning of hedges hard.

Core Claim

Assume the true one-period marginal distribution of the underlying moves by at most $\delta$ in 1-Wasserstein distance per period. Assume also that the empirical marginal from $w$ samples converges at the rate $C w^{-1/2}$ proved for the distributionally robust martingale transport problem in Distributionally Robust Martingale Optimal Transport. Then the empirical marginal built from the most recent $w$ observations lies, with high probability, within

$$\varepsilon(w) = C\,w^{-1/2} + \tfrac{1}{2}\,\delta\, w$$

of the current true marginal. The second term bounds the average displacement of a linearly drifting window. Minimizing over $w$ gives

$$w^\star = (C/\delta)^{2/3}, \qquad \varepsilon^\star = \tfrac{3}{2}\, C^{2/3}\delta^{1/3}.$$

The form comes from the same bias–variance balance used in nonstationary estimation, with the $1/2$ exponent inherited from the sampling rate.

For a payoff that changes by at most $L$ when the price moves one unit (an $L$-Lipschitz payoff), the robust price at radius $\varepsilon^\star$ is a valid superhedge for the current regime with controlled probability, and its excess cost over the true robust price is at most order $L\varepsilon^\star$.

Let $\varepsilon_{\text{crit}}$ be the radius at which the Wasserstein ball first contains a martingale law attaining the classical model-free bound. The bound is then informative exactly when $\delta < \delta_{\text{crit}} \propto \varepsilon_{\text{crit}}^3/C^2$. Two predictions are new and checkable. The first is a cube-root relation between drift rate and the robustness the data can support. The second is a cubic phase boundary, beyond which robust pricing has nothing to say.

Relation to Prior Work

The method builds directly on two Wasserstein-ball approaches to robust pricing. Distributionally Robust Martingale Optimal Transport relaxes the requirement that martingale transport match observed marginals exactly. It replaces exact matching with an $\varepsilon$-Wasserstein neighbourhood, proves strong duality between worst-case expectation and superhedging cost, and gives the $O(n^{-1/2})$ rate we reuse. Its radius, though, is static and its samples are assumed i.i.d. Robust Estimation of Superhedging Prices shows that plug-in superhedging estimators are fragile. It regularizes them with a Wasserstein ball around historical returns, with tests on GARCH simulations and S&P 500 and DAX data. It also works on stationary blocks and does not tie the radius to drift. Our contribution is the drift term and the threshold it implies. We do not touch the duality theory itself.

The streaming side supplies the measurement machinery and the main competitors. Beyond Euclidean Summaries tracks distributions in 2-Wasserstein space to raise change-point alarms but never prices anything. Fast Unsupervised Online Drift Detection Using Incremental Kolmogorov-Smirnov Test maintains empirical distribution functions with a randomized balanced search tree, at $O(\log N)$ cost per update. Here the high-dimensional cost of optimal transport, flagged as an open problem for these detectors, does not arise. Each one-period marginal is univariate, and the 1-Wasserstein distance between univariate empirical distributions is the area between their distribution functions. That area can be maintained with the same tree structure that powers that test.

Hedging the Drift adapts window size with an outer EXP3 bandit and a variation budget $B_T$. Our per-period $\delta$ is a Wasserstein analogue of that budget, used to set a radius in closed form rather than learned from rewards.

The PDE literature is a lens here, not a source of technique. Solutions in Morrey spaces of some semilinear heat equations with time-dependent external forces shows that a persistent, non-decaying forcing is tolerable as long as its norm stays below a threshold. Above that threshold, the guarantees fail. We look for the same structure in robust pricing, with drift playing the role of forcing. We do not claim any PDE estimate transfers. Duality formulas for robust pricing and hedging in discrete time narrows spreads by replacing superhedging with acceptance sets. That approach is complementary: our radius rule could feed such a formulation, but we keep strict superhedging so that coverage is easy to state.

Research Program

The theoretical work has three parts. We first prove the coverage statement above for drift processes with bounded per-period Wasserstein increments. This includes a version where $\delta$ is itself estimated from the stream, which requires a confidence bound on the drift estimate and adds a lower-order term. We then characterize $\varepsilon_{\text{crit}}$ for a few payoff families with tractable dual problems: forward-start straddles, and the VIX-type payoff studied in Bounds for VIX Futures Given S&P 500 Smiles. Finally, we state how heavy tails, which slow the sampling rate below $n^{-1/2}$, change the exponents. This is where the cube law should fail first, so it is a built-in falsification route.

The algorithmic work is an online estimator. It maintains empirical marginals over nested windows and estimates $\delta$ as the slope of 1-Wasserstein distance against lag between windows. It sets $w^\star$ and $\varepsilon^\star$ in closed form and re-solves the DRMOT linear program on a coarse grid. Abrupt regime changes violate the bounded-drift assumption. To handle them, we add a change-point trigger from the incremental test: an alarm resets the window and temporarily inflates $\delta$. The linear program's $O(n^{3d/2})$ scaling limits us to short horizons ($d\le 3$). We state this restriction up front and do not try to solve it.

Baselines

The naive reference is the sliding-window plug-in estimator. It applies martingale transport or superhedging directly to the last $w$ observations with no ball. Distributionally Robust Martingale Optimal Transport shows it is often infeasible, because empirical marginals violate convex order, and Robust Estimation of Superhedging Prices shows it is unstable. We run it with the window chosen by backtest and report infeasibility rate alongside shortfall.

The strongest existing method is static-radius DRMOT. It uses the radius rule from Distributionally Robust Martingale Optimal Transport, sized for sampling error only. It runs either on all history or on a backtest-tuned sliding window, with the window and radius tuned by grid search on a held-out early segment of each stream. That gives it a strictly larger tuning budget than our closed-form rule. Beating it means equal or lower shortfall frequency at matched or narrower price width on drifting segments, with no loss on stationary segments.

For pricing tightness, the external reference is Bounds for VIX Futures Given S&P 500 Smiles. On the May 5, 2016 smile it reports that the VIX-futures lower bound rises from the classical 0% to 8.4%, with the upper bound at 16.7%. Under a SABR model with 2-month maturity it reports a lower bound of 7.2% from the full linear program, about 5.2% from functionally generated portfolios, and an upper bound of 22.8%. For superhedging estimation on S&P 500, DAX 30, Apple and Gold returns, no reported numerical result was found on Lacuna. We will therefore re-implement the Obłój–Wiesel estimator on these series rather than quote a figure.

The adaptive-learning competitor is a Bandit-over-Bandit radius tuner in the style of Hedging the Drift. An EXP3 learner selects the radius from the same grid, rewarded by realized hedge outcome at maturity. That paper reports at least 24% less dynamic regret than obliviously tuned SW-UCB on CPRM-12-001, so it is a credible adaptive opponent.

The ablation that isolates the key ingredient keeps our window but removes the drift term, and separately keeps the drift term but fixes the window. If the sampling-only radius on the optimal window performs as well as the full rule, then the drift term is not doing the work.

Why not just…?

Why not just inflate a static radius by a safety factor chosen by backtest? This is the most serious threat. On a stream whose drift rate is roughly constant, a well-tuned constant radius will match ours. The proposal's claim is specifically about streams where drift speed changes. Our rule predicts that the right radius scales as $\delta^{1/3}$, so any constant is too wide in calm periods or too narrow in fast ones. The synthetic experiment varies $\delta$ in stages to test exactly this. If a single backtested radius stays within noise of ours across stages, we will report that the drift term is unnecessary in practice.

Why not just learn the radius from hedge outcomes with an adaptive bandit? The approach of Hedging the Drift assumes rewards arrive each round. A superhedge's outcome arrives only at maturity, so feedback is delayed by the option's life. Shortfalls are also rare by construction, so the learner gets little signal. Our rule uses unlabeled price observations, in the spirit of the label-free detection in Fast Unsupervised Online Drift Detection Using Incremental Kolmogorov-Smirnov Test. We run the bandit tuner head to head at several maturities. Our prediction is that its disadvantage grows with maturity. If it does not, the delayed-feedback argument is wrong.

Why not just reset on change-point alarms and keep a static radius between them? Pairing Beyond Euclidean Summaries or incremental Kolmogorov–Smirnov alarms with a static radius handles abrupt regime changes well. It is blind to gradual drift that never crosses an alarm threshold, and during the detection delay the hedge is misaligned. We test this on the paired Insects and Abrupt Insects streams. These share a sensor but differ in drift type, so the comparison should separate the regimes. We expect reset-only to win or tie on abrupt drift and lose on gradual drift, which is why our estimator includes the trigger rather than competing with it.

Proposed Experiments

The first experiment is synthetic. We simulate GARCH and SABR paths whose parameters drift at piecewise-constant rates spanning two orders of magnitude, interleaved with abrupt jumps. On each segment we measure three things: shortfall frequency (payoff exceeding the robust price), mean shortfall size, and price width relative to classical model-free bounds.

The decisive check comes from the same simulations. On log–log axes, the empirically best radius against true $\delta$ should have slope near $1/3$. The drift rate at which the robust lower bound for VIX-type payoffs falls to the classical value should scale as predicted by $\delta_{\text{crit}}\propto\varepsilon_{\text{crit}}^3/C^2$. A slope near $1/2$ or $1$ would falsify the mechanism. So would a soft rather than sharp degradation.

The second experiment uses the S&P 500, DAX 30, Apple and Gold daily return series from Robust Estimation of Superhedging Prices. We estimate rolling superhedging prices for short-dated Lipschitz payoffs and compare shortfall and width against all baselines, separately for calm periods and stress periods. As an anchor, the VIX experiment of Bounds for VIX Futures Given S&P 500 Smiles is reproduced on its SABR configuration, to confirm our linear-program implementation recovers the reported 7.2% lower bound before any drift is introduced.

The third experiment is a calibration test outside finance, using univariate features from the Bike, Insects and Abrupt Insects streams of the incremental Kolmogorov–Smirnov suite. Here we ask only whether the $\varepsilon^\star$ ball computed at time $t$ contains the empirical distribution of the next batch at the nominal rate. This checks that the coverage claim holds on real non-financial drift.

The project succeeds if three things hold together: the cube-root and cubic-threshold predictions hold in simulation, the full rule beats tuned static DRMOT on drifting segments of real returns without losing on calm ones, and the ablation shows the gain comes from the drift term.