Empirical martingale projections via the adapted Wasserstein distance

José Blanchet, Johannes Wiesel, Erica Zhang, Zhenyuan Zhang

This paper, recently submitted to the Annals of Applied Probability, introduces a robust mathematical framework for determining whether a dataset follows a "martingale" property—a cornerstone of financial mathematics and probability theory. By combining optimal transport theory with kernel smoothing, the authors provide a way to test for the absence of arbitrage in complex financial models.

TL;DR

The authors develop the Smoothed Empirical Martingale Projection Distance (SE-MPD), a new metric that measures how far a set of data points is from being a martingale. Unlike standard distance measures, this approach remains accurate in high dimensions and provides a consistent statistical test to detect hidden flaws (like arbitrage) in generative AI models used for asset pricing.

What Problem Does This Paper Address?

In finance, a "martingale" describes a process where the best prediction for a future value is its current value ($E[Y|X] = X$). This property is essential for "no-arbitrage" conditions—the idea that you shouldn't be able to make a risk-free profit.

Testing this property in real-world data is notoriously difficult for two reasons:

  1. The Continuity Gap: Standard ways of measuring distances between distributions (like the Wasserstein distance) are "blind" to conditional expectations. A tiny change in a distribution can cause a massive jump in its conditional expectation, making standard tools unreliable for martingale testing.
  2. The Empirical Problem: Raw data is discrete. If you have a single observation $Y$ for every $X$, the conditional expectation is just that single point, not a distribution. This makes raw empirical data look "non-martingale" by default.

Methods

The authors bridge these gaps using three key innovations:

  • Adapted Wasserstein Distance ($AW_\gamma$): Instead of using standard optimal transport, they use "adapted" (or bi-causal) transport. This version respects the flow of information over time, making it the "correct" geometry for studying martingales.
  • Kernel Smoothing: To fix the "empirical problem," they convolve the data with a smoothing kernel (specifically heavy-tailed densities). They prove a vital "invariance" property: if the original data was a martingale, the smoothed version remains one, and vice versa.
  • Closed-Form Characterization: They derive an explicit formula for the projection distance, making it computationally feasible to calculate how far a dataset is from the "space of martingales."
  • Asymptotic Theory: They prove that their test statistic converges at a parametric rate of $O(n^{-1/2})$. Crucially, this rate does not degrade as the number of dimensions increases, effectively "breaking the curse of dimensionality" that plagues most optimal transport methods.

Datasets

The paper validates the method using several types of data:

  • Synthetic Martingales: Random walks and processes constructed using Hermite polynomials to test the sensitivity of the "martingality" detection.
  • Markov Chains: Testing whether an ergodic sequence follows a specific transition kernel.
  • Generative Financial Models: They apply the test to Neural SDE-based European option calibration models. These are modern AI models used to price options, and the authors use their test to check if these models accidentally introduce arbitrage opportunities.

Key Insights and Ideas

  • The "Right" Topology: The paper reinforces that for any problem involving conditional expectations (like finance or reinforcement learning), the Adapted Wasserstein distance is superior to the standard Wasserstein distance.
  • Smoothing as a Feature, Not a Bug: Usually, smoothing is seen as a way to "blur" data. Here, it is a mathematical necessity that allows the discrete empirical measure to be projected onto a continuous space of martingale couplings without losing the core property being tested.
  • Dimension Independence: One of the most striking findings is that the SE-MPD converges at the same rate regardless of the dimension $d$. This makes it a powerful tool for high-dimensional financial portfolios.

Key Results

  • Closed-Form Formula: The authors show that the Martingale Projection Distance (MPD) can be simplified to a surprisingly elegant expectation: $2^{1-\gamma}E[|X - E[Y|X]|^\gamma_2]$.
  • Consistency: The proposed hypothesis test is shown to be "consistent," meaning that as you collect more data, the probability of correctly identifying a non-martingale process approaches 100%.
  • Detecting Arbitrage in AI: In their experiments on Neural SDEs, the authors found that some existing calibration methods—which appeared to fit market data well—actually failed the martingale test. This means those AI models were mathematically inconsistent and allowed for "free lunches" (arbitrage).

Limitations and Future Work

  • Bandwidth Selection: The performance of the test depends on the smoothing parameter $\sigma$. While the authors suggest a heuristic (maximizing the distance to increase test power), a rigorous "optimal" selection rule for $\sigma$ remains an open question.
  • Kernel Constraints: The current theory relies on specific types of heavy-tailed kernels. While they show Gaussian kernels can work under certain conditions, more work is needed to generalize the choice of smoothing functions.
  • Multi-period Extension: This paper focuses on pairs $(X, Y)$. Extending this to a full time-series (multi-period martingales) would be the next logical step for complex path-dependent financial derivatives.

Summary updated

Abstract

Given a collection of multidimensional pairs $\{(X_i,Y_i):1 \leq i\leq n\}$, we study the problem of projecting the associated suitably smoothed empirical measure onto the space of martingale couplings (i.e. distributions satisfying $\mathbb{E}[Y|X]=X$) using the adapted Wasserstein distance. We call the resulting distance the smoothed empirical martingale projection distance (SE-MPD), for which we obtain an explicit characterization. We also show that the space of martingale couplings remains invariant under the smoothing operation. We study the asymptotic limit of the SE-MPD, which converges at a parametric rate as the sample size increases if the pairs are either i.i.d. or satisfy appropriate mixing assumptions. Additional finite-sample results are also investigated. Using these results, we introduce a novel consistent martingale coupling hypothesis test, which we apply to test the existence of arbitrage opportunities in recently introduced neural network-based generative models for asset pricing calibration.
BibTeX
@article{blanchet2026empirical,
  author = {José Blanchet and Johannes Wiesel and Erica Zhang and Zhenyuan Zhang},
  title = {{Empirical martingale projections via the adapted Wasserstein distance}},
  journal = {The Annals of Applied Probability},
  year = {2026},
  doi = {10.1214/25-aap2239},
}
Figure