Mechanism-Aware Robust Pricing on Time-Shuffled Price Data

Privacy tools for streaming data increasingly leave the values alone and scramble when they occurred. Stateful Switch, for example, swaps each price with a neighbour inside a short window of $k$ slots, so every number is released exactly once and nothing is lost or duplicated. These mechanisms have been tested on stock prices, but only by asking whether a moving average survives. A derivatives desk asks a different question: what does a contract written on this series cost to hedge in the worst case? Reordering a price path manufactures reversals that never happened. It inflates realized variance and puts future information at past timestamps. A hedging analysis that takes the released series at face value can therefore report bounds that are wrong in either direction.

This note proposes to treat the privacy mechanism as part of the market model. The mechanism is a published algorithm whose randomness is secret, so its transition kernel is known. That means the set of true-market laws consistent with the released data can be written as a linear constraint. It can be added directly to the linear program used for martingale optimal transport. Martingale optimal transport is the computation of the cheapest portfolio that covers a contract's payoff under every arbitrage-free price law consistent with what is observed. The result is a robust price interval that is guaranteed to contain the interval computed from the unprivatized data, together with a measurement in dollars of what temporal privacy costs a hedger.

Core Claim

The robust pricing-hedging duality says that the cheapest superhedge (a trading strategy whose final value covers the payoff on every admissible scenario) equals the largest expected payoff over all martingale laws consistent with market information. In discrete time this sup is a linear program over path laws, as established in Duality formulas for robust pricing and hedging in discrete time and made computable by the ε-relaxed linear programs of Computational methods for martingale optimal transport problems. Let $P$ be the unknown law of the true discretized path, $K$ the known kernel mapping a true path to the law of its privatized release, and $\hat Q$ the empirical law of released paths. Then the mechanism-aware upper price is

$$\overline{V}_K = \sup \big\{ \mathbb{E}_P[g] \;:\; P \text{ is an } \varepsilon\text{-martingale law},\; \| K^{\top} P - \hat Q \| \le \delta \big\},$$

where $g$ is the payoff, $\varepsilon$ is the martingale-relaxation tolerance of Guo and Obłój, and $\delta$ is a sampling-error radius for $\hat Q$. The lower price is defined symmetrically. Because $K^\top P$ is linear in $P$, the problem remains a linear program. Its dual splits into the familiar static and dynamic hedging positions plus a new set of dual variables priced against the released data. If the true law lies in the constraint set, which holds with high probability when $\delta$ is chosen correctly, $[\underline V_K, \overline V_K]$ contains the true robust interval.

The substantive conjecture is about where the cost falls. Terminal-value payoffs such as European calls should be almost unaffected, because a window-$k$ swap only blurs the terminal marginal across $\pm k$ slots. Path-dependent payoffs such as variance swaps, lookbacks and Asian options should be badly distorted by naive plug-in, because swaps add artificial squared increments and fake extrema. On these payoffs only the mechanism-aware interval should remain both valid and informative.

Research Program

The work has three parts.

Analysis. For Stateful Switch (Stateful Switch: Optimized Time Series Release with Local Differential Privacy) and the Threshold Mechanism (Beyond Value Perturbation: Local Differential Privacy in the Temporal Setting), we will derive two things:

  • the exact kernel $K$ on a discretized grid of $n$ steps and $q$ price levels;
  • a simple bound on how much the released realized variance and running maximum can differ from the true ones as a function of $k$.

These bounds predict which payoffs need mechanism-awareness at all.

Solver. We will add the $K^\top P$ constraint to the Guo–Obłój linear program and read off the dual. The new dual variables are interpretable as "how much the hedger would pay to see unshuffled data." That price of privacy in dollars is the quantity the existing privacy literature lacks.

Mechanism comparison. Among switch rules with the same temporal privacy guarantee, we will measure which ones inflate the robust interval least. This turns hedging cost into a utility metric for choosing mechanisms, in place of moving-average error.

The closest prior work sets up but does not address this question. Ye et al.'s two papers build the mechanisms and evaluate them on Yahoo Finance daily prices only through simple-moving-average error; they impose no martingale structure, which is the gap here. Sequential Minimax Games as Stacked Martingale Optimal Transport decomposes robust hedging into one-step transport problems. We cannot reuse that stacking directly, because the switch kernel couples neighbouring steps, and we say so rather than assume it away. General Duality and Dual Attainment for Adapted Transport supplies the conceptual warning that strategies must respect the information flow. The released process at slot $t$ can depend on true information up to $t+k$, which explains why naive backtests on shuffled data can show phantom arbitrage. EHRCon motivates the stance: its finding that over 56% of note-table discrepancies in MIMIC-III are time-related, dominated by a one-hour offset, shows that de-identification artifacts are systematic consequences of a known pipeline rather than random noise. Systematic, known artifacts are exactly what a kernel constraint can model. We do not test clinical data here.

The scope limit is size. The path-law linear program has $q^n$ variables, so exact solutions are restricted to short horizons (around $n \le 6$, $q \le 15$, $k \in \{1,2,3\}$). For longer horizons we will use the random-discretization variant of Guo and Obłój. The file itself flags multi-step transport as hitting the curse of dimensionality, and we will report where the approach stops scaling.

Baselines

Naive plug-in. This is the "do nothing" reference: run the same Guo–Obłój linear program on the released paths as if they were true. It is the strongest practical competitor because it is what any analyst would do. Small-$k$ shuffles of daily data may barely matter for monthly terminal payoffs, and if so our method adds nothing there. It uses the identical grid, $\varepsilon$, and solver. We beat it if the plug-in interval fails to contain the true-data interval, or is shifted, on path-dependent payoffs while ours stays valid.

Unprivatized oracle. The same linear program on the original series gives the target interval. Every method is scored against it.

Generic relaxation ablation. This isolates our key ingredient. We drop the kernel constraint and instead loosen the martingale tolerance $\varepsilon$ until the plug-in interval covers the oracle. Covering is the same validity requirement our method guarantees, so this compares how much each approach must widen to achieve it. It tests whether knowing the mechanism matters or any uniform widening suffices. We beat it if, at matched coverage, our intervals are narrower.

Neural SDE. The strongest learned competitor is the martingale-constrained neural SDE of Robust pricing and hedging via neural SDEs, calibrated on released data with its published Augmented Lagrangian procedure. That paper reports calibration MSE of $10^{-8}$ to $10^{-9}$ on its Heston-generated surface. No reported result was found on Lacuna for its bounds on privatized data. It receives the same data and a matched wall-clock budget.

Value-perturbation release. Laplace noise on prices at matched privacy budget, via the factor-of-two equivalence Ye et al. prove between value and temporal privacy. On the privacy side, the strongest number in the file is that Stateful Switch's moving-average MSE is 3 to 5 orders of magnitude below value perturbation on Yahoo Finance stocks. No reported result was found on Lacuna for any mechanism's effect on robust option bounds.

Why not just…?

Why not just run robust pricing on the released data directly? This is the real threat. For European payoffs at horizons much longer than $k$ we expect it to work nearly as well, and we will report that honestly. The experiments decide where it breaks. Our prediction is that artificial reversals push realized variance, and therefore variance-swap bounds, upward, and inflate lookback bounds, by amounts that grow with $k$. If plug-in error stays inside sampling noise for every payoff class, the method is unnecessary and the main finding becomes "temporal privacy is hedging-safe," which is still worth publishing.

Why not just widen the martingale tolerance until the bounds are safe? Uniform widening ignores that the mechanism's damage is structured: it touches only $\pm k$ neighbourhoods and leaves the terminal marginal almost intact. The matched-coverage ablation measures this directly. If uniform widening is as tight as ours, the kernel constraint adds no value and we will say so.

Why not just release with value noise and keep timestamps exact? Exact order preserves the martingale property in expectation. But Ye et al. report value noise is orders of magnitude worse on moving averages, and it drives taxi trajectory clustering to near-zero normalized mutual information against 0.6–0.7 for temporal mechanisms. We will add its bound widths to the comparison so the trade-off is shown in hedging units too.

Proposed Experiments

Synthetic Heston paths. These give a known martingale law, so the true robust interval is computable exactly. We will apply Stateful Switch and the Threshold Mechanism over a grid of $k$ and privacy budgets. For each method we report:

  • coverage of the oracle interval across repeated releases;
  • width inflation relative to the oracle;
  • bias of the interval midpoint, for European calls, variance swaps, lookbacks and Asian payoffs.

Real data. We will repeat this on the Yahoo Finance daily US stock prices used by Ye et al., pooling short windows across stocks to estimate $\hat Q$. The real-data oracle is the bound computed on unshuffled windows. As a coarse-time check we will also use the monthly S&P 500 series from Reflexivity as Prompt. That paper already strips dates, so it tests whether date removal alone, with order intact, harms bounds.

Decisive outcomes. The proposal succeeds if two things hold:

  1. Plug-in coverage drops well below nominal for path-dependent payoffs as $k$ grows while ours stays near nominal.
  2. At matched coverage, ours is strictly narrower than uniform ε-widening.

It fails cleanly if plug-in remains valid throughout, or if uniform widening matches us.

A secondary check backtests the dual hedging strategy on released data. It confirms that phantom arbitrage from look-ahead appears under plug-in and vanishes under the mechanism-aware dual.