Measuring the Speed Demons: Hasbrouck and Saar on Low-Latency Trading

Around 2010, high-frequency trading became the villain of finance. It was blamed for the Flash Crash, accused of front-running ordinary investors, and denounced in bestselling books. Regulators wanted to act.

There was an awkward problem underneath all this noise. Nobody could reliably measure how much high-frequency trading was even happening. Orders do not arrive at the exchange with a label saying "sent by a high-frequency firm." Exchanges know who their members are, but researchers do not, and the handful of studies that had privileged access to identified data could not be replicated by anyone else.

Joel Hasbrouck and Gideon Saar solved this with a genuinely clever piece of detective work, and then used their new measuring stick to ask the question everyone was arguing about.

The problem: you cannot regulate what you cannot measure

To study the effects of high-frequency trading you need to know, day by day and stock by stock, how much of it there is. Without a measure, every claim about HFT, that it helps liquidity, that it hurts it, that it causes volatility, is just assertion.

The gold-standard datasets, where an exchange tells you which orders came from which type of firm, exist but are rare and restricted. What everyone does have access to is the raw message data: the complete, timestamped stream of every order submission, cancellation and modification hitting the exchange, with no names attached.

The question is whether you can spot the fast traders in that anonymous stream by their behaviour alone.

The key idea via analogy: identifying a hummingbird by its wingbeat

You cannot see who is in the room. But you can hear them.

If you record the sound of a room full of birds, you can tell there is a hummingbird present without ever seeing it, because nothing else beats its wings at eighty times a second. The signature is in the timing.

Hasbrouck and Saar's insight is the same. A human trader, or even a slow algorithm, operates on a timescale of seconds. A high-frequency algorithm operates on a timescale of milliseconds. So if you look at the message stream and find chains of activity where an order is submitted and then cancelled and then replaced faster than any human could possibly react, you have found an algorithm running at machine speed.

They formalise this by building what they call strategic runs: linked sequences of order submissions, cancellations and executions that clearly belong to a single participant working a single strategy, identified by the fact that they happen in tight, machine-speed succession. Count the runs that occur within very short time windows and you have a proxy for low-latency activity in that stock on that day.

Then they did the essential validation step. They compared their behavioural measure against the NASDAQ-constructed estimates of high-frequency trading, the ones built from actual firm identities. The two are highly correlated. Their detective-work proxy tracks the real thing.

That is the methodological contribution, and it is a big one, because it means any researcher with ordinary message data can now measure low-latency activity. The measuring stick was democratised.

What they found

Armed with the measure, they asked whether more low-latency activity makes markets better or worse for everyone else. The results, at least in the period and market they studied, cut against the popular narrative.

Increased low-latency activity was associated with improvements in traditional market quality measures.

• Spreads got narrower. The gap between the best bid and the best ask, which is what an ordinary investor pays to trade, shrank.
• Displayed depth increased. There was more size resting in the order book, meaning larger orders could be absorbed.
• Short-term volatility fell. Prices jittered around less on very short horizons.

In short, on the measures that regulators traditionally use to judge whether a market is healthy, more speed looked like a good thing, not a bad one. The fast traders appeared to be doing what market makers are supposed to do: tightening quotes, adding depth, and damping noise.

Why it mattered

• It gave the whole field a measuring stick. This is arguably the paper's largest contribution. By showing that low-latency activity can be inferred from public message data, and validating that inference against identified data, it made an entire literature possible.
• It supplied evidence into an evidence-free debate. The high-frequency trading argument was being conducted almost entirely on anecdote and intuition. This was careful, credible empirical work on a question of real public importance.
• It complicated the villain narrative. The finding that speed coincided with tighter spreads and deeper books made it much harder to sustain the simple story that high-frequency trading is pure predation.
• Strategic runs are a useful object in their own right. The idea of reconstructing a single participant's strategy from an anonymous message stream by chaining together machine-speed events has been reused widely.

The honest limitations

• Association is not causation, and the authors know it. Days with more low-latency activity are also days when other things are different. Establishing that speed causes the improvement, rather than merely accompanying it, is very hard. The companion literature, notably Hendershott, Jones and Menkveld, went hunting for genuine natural experiments precisely because of this problem.
• Averages hide the tails. The result is that market quality is better on average. It says nothing about what happens on the rare, catastrophic days when the fast traders all withdraw at once. The Flash Crash is exactly a story about the tail, and an average-based study is structurally unable to speak to it.
• The traditional measures may be measuring the wrong thing. Narrow quoted spreads look great in a regression. But if the displayed size is tiny and vanishes the instant a real order arrives, the quoted spread flatters the market. Whether a large institution genuinely gets a better fill is a different and harder question.
• The proxy is a proxy. Strategic runs capture low-latency activity, which is not the same thing as high-frequency trading. A slow institution using a fast execution algorithm will show up in the measure. The correlation with identified data is high but not perfect.
• One market, one era, one asset class. The study covers NASDAQ equities in a specific period. Speed's effects may differ elsewhere, and the arms race has escalated considerably since.

The one-line takeaway

Hasbrouck and Saar found a way to identify high-frequency traders by their machine-speed fingerprints in anonymous message data, validated it against real identified data, and then used it to show that in the market they studied, more speed came with narrower spreads, deeper books and lower short-term volatility.