Anomaly Detection: VolGAN & the Obsidian Screener

Most of this institute has trained you to ask one question: when a number moves, is that move genuine evidence the asset will do something, or is it noise dressed up as a signal? Anomaly detection is that same question turned inside out. Instead of asking whether a move is large, you ask whether it is strange — whether the market has just produced a state it almost never produces. And the deep and subtle point is that those two questions are not the same. A ten percent move can be utterly ordinary for a meme stock and a screaming anomaly for a sovereign bond. Size is relative to a distribution. Strangeness is the distribution.

So let us define anomaly the way a machine learning system actually defines it, because the intuitive definition — "something unusual" — is too vague to compute.

An anomaly is a point that the model assigns low probability. You learn, from a great deal of normal history, a probability distribution over market states. Then a new state arrives, you evaluate how likely that state was under what you learned, and if the likelihood is small enough, you flag it. That is the whole idea. Everything else — the neural networks, the adversarial training, the screeners — is machinery for estimating that distribution when the state is too high-dimensional to model by hand. And the volatility surface is exactly such a state: a whole grid of implied volatilities across strikes and maturities, dozens of numbers that must move together in tightly constrained ways. You cannot eyeball whether today's surface is plausible. You need a model that has seen thousands of plausible surfaces and learned their shape.

This is where the generative adversarial network earns its place. In SVQ, that model is called VolGAN, and it is trained to do one deceptively simple thing: generate volatility surfaces that are indistinguishable from real ones.

Here is the intuition before the formalism. A GAN is a contest between two networks. The generator takes random noise and tries to manufacture a fake volatility surface. The discriminator is shown a mixture of real historical surfaces and the generator's fakes, and its only job is to tell them apart. They train against each other. Every time the discriminator gets sharper at spotting fakes, the generator is forced to make better fakes; every time the generator improves, the discriminator must find subtler tells. Formally, the two play a minimax game over a value function — the generator minimizing, the discriminator maximizing, the expected log-probability that the discriminator correctly classifies real versus generated samples. At the equilibrium of that game, the generator's output distribution has converged onto the true distribution of volatility surfaces, and the discriminator is reduced to a coin flip, unable to tell real from fake because there is no longer a difference to find.

Now sit with what you have built. After training, the generator is a learned model of every normal volatility surface the market produces. It knows that the smile curves a certain way, that the term structure slopes a certain way, that the wings cannot detach from the body without the whole thing looking wrong. That knowledge is precisely what you need to detect an anomaly, and there are two clean ways to extract it.

The first uses the discriminator directly. Feed it today's real surface. If the discriminator — trained to recognize authentic surfaces — judges today's surface as improbably fake-looking, that is a signal the surface has drifted off the manifold of normal states. The second, and often more robust, approach is reconstruction. You ask: what is the closest surface the generator can produce to today's observed surface? You search the noise input for the latent point whose generated surface best matches reality, and you measure the leftover gap. For a normal surface the generator can reproduce it almost exactly and the gap is tiny. For an anomalous surface — one the generator has effectively never been asked to make — the best it can do still leaves a large residual. That residual, the reconstruction error, is your anomaly score. Large residual, low likelihood, strange state.

A score by itself decides nothing. You need a threshold, and that is what the Obsidian screener supplies. Obsidian runs VolGAN's anomaly score across the universe continuously and surfaces only the names whose score crosses a calibrated cutoff — turning a continuous strangeness measure into a discrete watchlist of surfaces that have genuinely left their normal envelope. The cutoff is usually set as a high quantile of the historical score distribution, so that on a quiet day almost nothing fires and you are not drowned in false positives.

And here the institute's throughline returns with teeth. A high anomaly score is not the same as evidence the asset will move. It is evidence the asset is in a state it rarely occupies — which is a different, and weaker, claim. The discipline is to separate three things that all produce a high score: a true structural dislocation that often does precede a move; a data artifact, like a stale quote or a thin-strike surface that VolGAN was never trained to handle; and ordinary regime drift, where the whole market has shifted and every surface looks novel relative to a model trained on the old regime. Only the first is tradable. The screener hands you the candidates. Judging which kind you are looking at is the quant's job, and it is the entire reason you are being trained rather than automated.