One Law for All Stocks: Sirignano and Cont Find a Universal Price Formation Model

Every stock is different. A megacap technology name trades hundreds of millions of shares a day, in penny increments, with a spread that is essentially always one tick. An illiquid small-cap trades in fits and starts, with a wide spread and a thin book. They have different tick sizes, different participants, different rhythms.

So the natural way to model them is separately. Fit a model to each stock, tuned to its own particular behaviour. That is what practitioners generally did, and it is the obvious thing to do.

Justin Sirignano and Rama Cont tested that assumption against a mountain of data, and found it was wrong.

The problem: is price formation stock-specific or is it a mechanism?

Here is the question, stated cleanly.

When an order arrives at the book, the price responds. Is the rule governing that response a property of the individual stock, learned separately for each one? Or is there a common mechanism that operates the same way everywhere, with the differences between stocks being surface details rather than different physics?

This is not an idle question. It has an enormous practical consequence. If price formation is stock-specific, you must fit a separate model per stock, and for a stock with thin data you will never have enough observations to fit anything reliable. If price formation is universal, you can pool every stock's data into one gigantic training set, and a stock with a sparse history inherits everything learned from the liquid ones.

To answer it, you need a model flexible enough that it cannot be accused of imposing the answer through its own assumptions, and you need a truly enormous amount of data.

The key idea via analogy: the language of the order book

Sirignano and Cont assembled a high-frequency dataset covering billions of market events (orders, cancellations and trades) across a broad set of US equities, and trained deep neural networks to predict the next price move from the recent history of the order book.

Then they ran the decisive experiment. They trained two kinds of model.

The stock-specific models: one network per stock, trained only on that stock's own data.

The universal model: a single network, trained on all the stocks pooled together, with no stock identity given to it at all. It does not know which stock it is looking at. It just sees an order book and its recent history.

The naive expectation is that the specialists win. Each one is tuned to its own instrument, learning its idiosyncrasies. The generalist is a compromise.

The generalist won. The single universal model, trained on all stocks together, produced better out-of-sample predictions than the models trained individually on each stock. And, most tellingly of all, it worked on stocks that were not in its training data at all.

The analogy is to language. You could try to learn each speaker's speech patterns separately, one model per person. Or you could realise that everyone is speaking English, learn the grammar of the language from all speakers pooled together, and then understand a speaker you have never met. Individual speakers have accents and quirks, but the grammar is shared, and the grammar is where the information is.

Sirignano and Cont's finding is that the order book has a grammar. The dynamics of supply and demand map to subsequent price moves through a mechanism that is shared across instruments, and once you have learned that mechanism from enough data, the per-stock quirks matter less than the common structure.

Why the pooling wins, and why it is not a trick

There are two reasons the universal model beats the specialists, and both are worth understanding.

First, data. Any individual stock, even a heavily traded one, gives you a limited sample. Pooling across the whole market gives you orders of magnitude more. A deep network with many parameters is starved by the former and thrives on the latter. More data means less overfitting and a genuinely better estimate of the underlying relationship.

Second, and more fundamentally, the thing being learned is genuinely common. If price formation really were stock-specific, then pooling would be an error, mixing together different mechanisms and averaging them into mush. More data would not rescue a model that was learning a nonexistent common law. The fact that pooling helps, and that it transfers to unseen stocks, is evidence that the common law exists.

That second point is the scientific content of the paper, and it is why it is not merely an engineering result. The authors also found that the relationship is non-linear and depends on the state of the book across multiple levels and on its recent history, which explains why the classical linear microstructure models, elegant as they are, were leaving predictive information on the table.

There is a further striking finding: the model exhibits stability over time, continuing to perform out of sample across periods well after its training data. That is unusual and important in a field where signals typically decay within weeks.

Why it mattered

• It is evidence for a real, shared mechanism. Most machine learning results in finance are "our model predicts better." This one makes a claim about how markets work, and it backs the claim with the strongest kind of evidence: a model trained on some instruments works on others it has never seen.
• It changed how quants build microstructure models. Pooling across instruments to train one model, rather than fitting per-instrument models, became a standard approach. For illiquid instruments this is transformative, because they can now borrow strength from the liquid ones.
• It arrived alongside DeepLOB, and they agree. Zhang, Zohren and Roberts found the same transferability from a different architecture, a different market, and a different dataset. Two independent teams reaching the same conclusion through different routes is far more persuasive than either result alone.
• It vindicated deep learning on market data, at scale. The paper made the case that the reason earlier neural network attempts on financial data had disappointed was not that markets are unlearnable, but that people were training on far too little data.

The honest limitations

• Prediction is not profit, and the gap is the whole game. The model predicts the direction of the next price move. At these horizons, the move you have predicted is often smaller than the bid-ask spread you would have to cross to capture it. High directional accuracy and zero trading profit are entirely compatible, and the paper does not build a trading strategy.
• Latency is unaddressed and may be fatal. A deep network takes real time to evaluate. In a domain where the competition is measured in microseconds, the time it takes to run your model may exceed the lifetime of the signal it produces. A universal law of price formation is not worth much if you cannot act on it in time.
• The data requirement is a moat. Billions of message-level events across a broad universe of stocks is not something a researcher can casually acquire. This result is, in a real sense, only reproducible by institutions with the data and the compute, which is uncomfortable for a scientific claim.
• "Universal" is a strong word for one market in one period. The evidence is from US equities over a particular window. Whether the same mechanism governs futures, foreign exchange, crypto, or US equities under a different market structure regime is an open question. Universality claims should always be treated as provisional.
• It is a black box making a scientific claim. The paper says a shared mechanism exists. It cannot tell you what that mechanism is, in a form a human could write down and reason about. That is genuinely unsatisfying, and it limits how much you can trust the model when conditions shift.
• The edge, if it is tradeable, will be competed away. Microstructure is adversarial. Any exploitable regularity in the order book is one that many well-resourced participants are hunting, and finding it changes it.

The one-line takeaway

Sirignano and Cont trained deep networks on billions of order book events and found that a single model trained on all stocks together outperforms models tuned to each stock individually, and works on stocks it has never seen, which is strong evidence that price formation follows a universal mechanism rather than a set of instrument-specific quirks.