How Much Does a Trade Actually Tell Us? Hasbrouck's Measure of Information Content

Theory said trades carry information. Kyle said informed traders hide their bets in the order flow. Glosten and Milgrom said the market maker learns from each trade and revises quotes accordingly.

Fine. But how much does a trade teach the market? Is a 10,000-share buy twice as informative as a 5,000-share buy? Does the market absorb the lesson instantly, or does it take an hour to sink in? Nobody had a good way to ask.

Hasbrouck's 1991 paper is the one that made the question answerable, and the answers turned out to be genuinely surprising.

The problem: a trade moves the price, but so does everything else

You want to know what a trade "did" to the price. Naively, look at the price before and the price after. Subtract.

This fails immediately, for two reasons.

First, prices move on their own. Some of the move after your trade is just news arriving that had nothing to do with you. You cannot attribute all of it to the trade.

Second, and more subtly, trades are not random events. They cluster, they follow each other, they respond to price moves as much as they cause them. If a price starts rising and traders chase it, you will see "buy, price up, buy, price up" and conclude that buying causes prices to rise, when the causation may partly run the other way. Trades and quotes are locked in a feedback loop, and you cannot cut it with a subtraction.

You need a method that can untangle a two-way conversation.

The key idea via analogy: the echo that never fades

Here is Hasbrouck's framing.

Think of the market as a room with an echo. Someone shouts (a trade arrives). The room responds: an immediate bang, then reverberations that decay over the following seconds and minutes. But the room is also full of other people shouting for their own reasons, and those shouts cause the first person to shout too. To measure the true effect of one shout, you must model the whole conversation and then ask: if I inject one surprise shout into this system, out of nowhere, what happens to the room over the next several minutes, and where does the sound level settle when everything calms down?

That is exactly what Hasbrouck did. He modelled trades and quote revisions jointly, as a system where each one depends on the recent history of both. Then he isolated the piece of a trade that could not have been predicted from anything that came before, the surprise in the order flow, sometimes called the trade innovation. That surprise is the only part that can carry genuinely new information, because the predictable part was already priced in.

And then the crucial move: he traced the surprise forward, and asked where the price ends up permanently. Not the immediate jump, which mixes real information with temporary pressure and bid-ask bounce, but the level the price settles at forever after. That permanent, never-reverted portion of the price move is the trade's information content. It is what the market genuinely learned.

What he found, and why it surprised people

Three findings from the paper are worth stating plainly, because they reshaped how practitioners think about execution.

Information arrives with a long lag. A trade's full price impact does not land instantly. It arrives over an extended period, with the price continuing to drift in the trade's direction well after the trade has printed. This is deeply counterintuitive if you believe markets are instantly efficient, and it is one of the most practically important facts in all of execution: your footprint keeps growing after you are done.

Impact is a concave function of size. Bigger trades are more informative, yes, but not proportionally so. Doubling the size does not double the permanent impact. This concavity, an early empirical cousin of what would later become the square-root law of market impact, means large trades are less costly per share than a naive linear model would predict, though still costly.

Information asymmetry is bigger in smaller firms. Trades in small-cap stocks carry proportionally more information than trades in large ones, which fits the intuition that fewer analysts and less disclosure means more room for someone to know something you do not.

He also found that large trades widen spreads, and that trades occurring when spreads are already wide have larger impacts, both signals that market makers are actively defending themselves against exactly the adverse selection the theory predicted.

Why it mattered

• It made adverse selection measurable per stock, per trade. Glosten and Harris had decomposed the spread. Hasbrouck went further and gave a full dynamic picture: not just how much of the spread is information, but how the information seeps into the price over time.
• It gave execution its central concept: permanent versus temporary impact. Every transaction cost model in production today, including Almgren and Chriss and its many descendants, rests on the distinction Hasbrouck operationalized. The permanent part is what you pay for revealing yourself. The temporary part is what you pay for being in a hurry. They must be modelled separately, and this is the paper that showed you how to see them separately in real data.
• It introduced the vector autoregression to microstructure and never left. The technique of modelling trades and quotes as a mutually dependent system, then reading off the long-run response to an order flow shock, became the field's standard econometric hammer. Hasbrouck himself would reuse it four years later for the information share.
• It quantified something traders felt but could not prove. Every desk knew that big orders "leave a mark." Hasbrouck told them how big the mark is, how long it takes to appear, and how it scales.

The honest limitations

• It depends on trade classification. The whole apparatus needs to know whether each trade was a buy or a sell, and that has to be inferred, typically with the Lee and Ready rule. Misclassified trades contaminate the estimates in ways that are not just noise.
• Permanence is defined by the model, not observed. "The price impact that never reverts" is read off the long-run behaviour of an estimated linear system. If the true dynamics are not linear, or if the model is misspecified, what looks permanent may not be. It is an inference, and a model-dependent one.
• Linearity is a strong assumption. The framework is fundamentally linear, and it captures the concavity of impact by allowing size to enter flexibly, but the deeper nonlinearities that later work (Lillo, Farmer, Bouchaud) documented sit uneasily inside it.
• Correlation is not causation, even here. The "surprise" in order flow is defined as the part unpredictable from past observed variables. If the informed trader and the price both respond to something the econometrician cannot see, the attribution of causality gets murky. This is an honest, and probably unfixable, limitation of the approach.
• The data was NYSE in the late 1980s. A world with one specialist, wide tick sizes and slow reporting. The methods survive; the magnitudes belong to that era.

The one-line takeaway

Hasbrouck showed how to measure what a trade actually teaches the market: model trades and quotes as one conversation, isolate the genuinely surprising part of the order flow, and follow it until the price stops moving, and what you find is that a trade's full information impact arrives slowly, grows less than proportionally with size, and is largest in the smallest firms.