Did the Robots Make Markets Better? Hendershott, Jones and Menkveld Find a Natural Experiment

By the late 2000s, algorithmic trading had gone from a curiosity to the majority of volume, and the public argument about it had become loud and evidence-free.

One side said: algorithms are efficient, tireless, and competitive. They tighten spreads and cut costs for everyone. The other side said: algorithms are parasites. They exploit speed advantages, they vanish when you need them, and they extract money from real investors.

The problem was that both sides could point at the same data and win. Algorithmic trading grew. Spreads narrowed. But so what? Lots of things were changing at once, and correlation is not causation.

Hendershott, Jones and Menkveld's 2011 paper is admired less for its finding than for how it got there.

The problem: you cannot run the experiment

The gold standard for a causal claim is a randomized experiment: take a market, randomly assign it more algorithmic trading, and see what happens.

You obviously cannot do that. So you are stuck with observational data, and observational data on this question is poisoned by a nasty problem called reverse causality.

Here is the trap. You observe that stocks with more algorithmic trading have tighter spreads, and you conclude algorithms tighten spreads. But think again: algorithms go where trading is easy. Algorithmic strategies naturally gravitate to stocks that are already liquid, because that is where their strategies work. So liquid stocks attract algorithms, rather than algorithms creating liquidity. The correlation is real, and the causal story could run entirely backwards.

Worse, some third thing, a change in the stock's investor base, a shift in volatility, an index inclusion, could be driving both. You are stuck.

The key idea via analogy: the accidental push

What you want is an exogenous shock: something that shoved algorithmic trading upward for reasons that had nothing to do with liquidity. If you can find a push that came from outside the system, you can watch what happened afterwards and legitimately call it causation.

The authors found one. In 2003, the NYSE began automatically disseminating its quotes through a system called Autoquote, rolling it out across stocks in stages rather than all at once. Before, quote updates were published manually by the specialist's clerk. After, they were pushed out automatically and continuously.

Why does this matter? Because algorithms need machine-readable, fast-updating quotes to function. An algorithm cannot trade against a quote it cannot see in time. Turning on automated quote dissemination was, in effect, opening the door for the machines.

And here is the crucial part: the NYSE did not turn on Autoquote in order to change liquidity. It was a technology upgrade, phased in across stocks for operational reasons that had nothing to do with each stock's liquidity. That makes it, for the researcher's purposes, close to a random push. It is a natural experiment.

So the design becomes: watch what happened to liquidity in a stock right after Autoquote arrived for that stock, compared to stocks where it had not arrived yet. Any change is attributable to the surge in algorithmic trading the technology enabled, and not to reverse causality, because the timing was imposed from outside.

What they found

The results, particularly for large stocks, were:

• Spreads narrowed. Algorithmic trading made it cheaper to trade.
• Adverse selection fell. This is the more subtle and more interesting result. The information content of trades declined, meaning liquidity providers were being picked off less. Algorithms appear to be better at avoiding informed counterparties, and that saving is passed on in tighter quotes.
• Quotes became more informative. More of the price discovery happened through quote updates rather than through trades. The market's posted prices were doing more of the work of telling you what the asset was worth, and trades were doing less of it.

Put together, the picture is: algorithms made liquidity provision better, not worse, at least on these measures, at least for large stocks, at least in 2003. Every qualifier in that sentence is load-bearing, and the authors were careful about all of them.

Why it mattered

• It is a methodological landmark. This paper is taught as a model of how to establish causality in finance when you cannot run an experiment. The hunt for an exogenous instrument, the staged rollout, the difference-in-differences logic: this is how the credibility revolution reached market microstructure.
• It gave the policy debate a real fact. Regulators worrying about algorithmic trading finally had a credible causal estimate rather than duelling anecdotes. The finding that algorithms improved liquidity was not what many critics expected, and it materially shaped the tone of subsequent regulation.
• It reframed what algorithms do. The finding that adverse selection fell suggests algorithms are not primarily predators picking off slow humans. They are, at least in part, better-defended liquidity providers, and being harder to pick off makes them willing to quote tighter. That is a genuinely different picture from the popular narrative.
• It launched the modern empirical HFT literature. Brogaard, Hendershott and Riordan on price discovery, Menkveld on the new market makers, and the whole subsequent wave of careful empirical work on fast trading are the children of this paper's approach.

The honest limitations

The authors themselves flag most of these, and it is one of the reasons the paper is so well regarded.

• The results are concentrated in large stocks. The evidence for small stocks is much weaker. Whatever algorithms did, they did it mostly for the names that were already easy to trade. This is not a small caveat: the stocks that most need liquidity got the least benefit.
• 2003 is not 2011, let alone today. Autoquote enabled algorithmic trading, but the algorithmic trading of 2003 was a far cry from the microsecond arms race that followed. It is a genuine leap to read this paper as a verdict on modern high-frequency trading, and the authors do not claim it is one. The mechanism they identified may not describe a world of colocation and latency arbitrage.
• Narrow measures of liquidity. Spreads and adverse selection are good measures for small trades. They say much less about whether a large institutional order can be executed without disaster, and there is a persistent complaint from the buy side that the "liquidity" algorithms provide is thin, fleeting, and evaporates precisely when you try to use it. This paper's measures cannot see that.
• It says nothing about stability. Improved average liquidity is compatible with catastrophic tail behaviour. The flash crash was three years ahead. A market that is cheaper in normal times and more fragile in bad times may or may not be a better market, and this study is silent on the tradeoff.
• The instrument is not perfect. Autoquote's staged rollout was not literally random, and one can worry about whether the NYSE sequenced it in ways correlated with something else.

The one-line takeaway

Hendershott, Jones and Menkveld used the NYSE's staged rollout of automated quotes as an accidental experiment that pushed algorithms into some stocks before others, and found that algorithmic trading causally narrowed spreads and reduced adverse selection, at least for large stocks in 2003, which is a real answer to a question that had previously been all argument and no evidence.