Teaching generative models where arbitrage lives

A generative model trained to produce option implied-volatility surfaces can be extremely good at its literal training objective — matching the statistical shape of real market surfaces — and still routinely output surfaces that no rational market could ever price, because they imply a static arbitrage: a way to lock in riskless profit just from the shape of the surface itself. This paper doesn't try to filter those violations out after the fact. It goes into the model's latent space and studies where the good and bad outputs actually live.

A margin for every latent code

The approach is direct. For a fixed, already-trained generator, every point in its latent space maps to some generated volatility surface. The authors assign each latent code a scalar margin, computed from the no-arbitrage conditions that surface would need to satisfy. Codes with a non-negative margin generate admissible, arbitrage-free surfaces; the set of all such codes is the admissible latent region.

Because this treats the generator purely as a map from latent variables to surfaces, the analysis isn't tied to one architecture — it applies to variational autoencoders, generative adversarial networks, or any other generative model with a deterministic realization map from latent code to output. That generality is what makes the framework useful as a diagnostic rather than a technique bolted onto one specific model family.

Finding — and using — the boundary

The mathematical core of the paper characterizes that admissible region's boundary as the set of zero-margin codes, and establishes the conditions under which a strictly admissible code stays admissible under small perturbations — in effect, how much you can nudge a good latent point before it tips into producing an arbitrage violation. For the regular parts of that boundary, the authors formulate a level-set equation whose local dynamics point toward the zero-margin set, which is what turns this from a static classification into something actionable: a latent code that currently generates an invalid surface can be moved, following that dynamic, back toward the admissible region — repaired, rather than simply discarded.

A surface can be statistically realistic and still be financially inadmissible — the two are not the same claim, and conflating them is the recurring failure mode in generative finance.

What the VAE experiments show

Testing on analytical cases recovers known boundaries correctly, which is the sanity check any geometric method like this needs to pass before it's trusted on anything harder. The more interesting result comes from experiments with a variational autoencoder trained on Heston-model-generated volatility surfaces: surfaces with similar reconstruction error — meaning the VAE judges them equally good by its own training loss — can correspond to different admissible regions. Reconstruction accuracy, the metric the model was actually optimized against, and financial admissibility, the property that actually matters for using the surface, are simply not the same axis. A model can be doing very well by the number it was trained on and still be scattering some outputs outside the region where they're usable.

Honest caveats

This is methodological and geometric work validated primarily on analytical benchmarks and one VAE trained on model-generated (Heston) surfaces, not on a broad sweep of architectures trained on real market data — how the approach behaves on a GAN or diffusion model trained on live quoted surfaces, across many underlyings and market regimes, is the natural next test. Computing and correcting margins for every latent code also adds a genuine computational step on top of ordinary generation, and the paper doesn't report the throughput cost of doing this at production scale.

Why it matters

The paper is one clean instance of a principle that shows up constantly in financial machine learning: realistic-looking is not the same claim as admissible, and a model optimized purely for the former will drift into the latter's violations without ever being told to. Embedding a hard constraint into the geometry of generation — rather than hoping the generator learns it implicitly, or rejecting bad samples after the fact — is a more structurally sound way to get compliant behaviour. For derivatives desks and volatility quants, the immediate uses are concrete: generate synthetic scenarios that are guaranteed arbitrage-free for stress testing, repair a generator's outputs where they drift outside the admissible set, and check whether a trained latent prior actually concentrates where the financially sensible surfaces live before trusting anything it produces downstream.