When Machines Trade With Machines: What Changes When Most of a Market's Participants Are Systems

A market is a conversation among its participants, and the participants have changed. For most of the history of organized exchange, the parties on either side of a trade were people: a specialist on a floor, a broker on a telephone, a portfolio manager deciding in the moment. Over the past two decades the share of order flow that originates in software rather than in a human decision has grown from a curiosity to a majority in most liquid electronic markets. The change is usually discussed as a matter of speed. We think that framing misses the more consequential point. When most of the participants in a market are systems, the market's behavior is set less by what any one participant intends and more by how their rules interact.

This piece looks at what changes when the crowd is made of machines. We describe the shift in the composition of order flow in stylized terms, show why the volume of messages has grown far faster than the volume of trades, trace the feedback loop through which one system's action becomes another system's input, and catalog the behaviors that loop can produce when it runs at machine speed. None of this is an argument that automation has made markets worse; on most days it has made them cheaper and more reliable to trade in.^[1] It is an argument that they have become a different kind of object, one whose failure modes are collective rather than individual, and that a systematic investor should model them that way.

The Composition of the Crowd

It helps to be specific about who the machines are, because they are not one thing. The oldest layer is systematic execution: algorithms that take an order a person has already decided to place and work it into the market over minutes or hours, slicing it into child orders and repricing them as conditions change. The second is algorithmic market making: systems that post continuous two-sided quotes across many instruments and manage the resulting inventory with rules for spread, size, and withdrawal. The third is model-driven trading: strategies whose signals are generated statistically and whose orders follow from those signals with little human involvement per trade, at horizons from minutes to weeks. The newest layer is what we call autonomous decision systems: models that adapt their own policies from experience and operate with human review at the level of parameters and limits rather than individual decisions. The boundaries between these layers blur in practice, but the order of their arrival is roughly right.

Figure 1 stylizes that arrival as a sequence of adoption curves. Each automated layer follows a logistic path from negligible to mature, with later layers starting later, and the discretionary share is whatever remains. The numbers are invented and the horizontal axis is an abstract clock rather than a calendar, but the shape is the point. Each layer of automation is built on top of the last, and the discretionary residue at the end is thin: a set of participants who still decide in the moment, trading in a market whose every other feature is mechanical.

Two consequences follow from that composition. The first is that prices are set at the margin, and the marginal participant is now almost always a machine: the quote you see was placed by a rule, and the order that moves it was generated by another rule. The second is less obvious and, we think, more important. Machines resemble one another more than people do. They are trained on the same public data, built from the same literature, constrained by similar risk limits, and tuned to similar objectives. A crowd of people is diverse in its reactions because attention, temperament, and information differ from person to person. A crowd of systems can be uniform in its reactions without any coordination at all. Homogeneity, not speed, is the property that most changes how the market behaves.

More Messages, Not More Trades

The most visible signature of automated participation is not the number of trades but the number of messages. A market maker quoting a few hundred instruments must revisit every quote whenever any of its inputs changes: a correlated instrument ticks, its inventory shifts, an index future moves. An execution algorithm working a parent order reprices its child orders continuously as the book moves. Almost none of this activity results in a trade. It is the housekeeping of systems that keep their exposure where they want it, and the housekeeping scales with the number of instruments and the frequency of input changes rather than with anyone's intention to transact.

Figure 2 shows the result in a stylized model. The trade rate grows modestly as electronic access widens; the quote update rate grows with it and with a quote-to-trade ratio that rises along a logistic path from single digits to the hundreds. The vertical axis is logarithmic because the gap between the two lines is the story: in a mature automated market, most of what the feed carries is quote updates that will be revised or withdrawn before anyone trades against them.

Three things change when the ratio is that high. The information content of any single message falls, so a participant that reacts to each one is mostly reacting to other participants' housekeeping. The displayed book becomes contingent: what is visible is a snapshot of many systems' current intentions, most of which will change before a marketable order can reach them. And the cost of understanding the market shifts from trading capacity to processing capacity, because the raw feed is now too fast for a person to read and must be summarized by yet another system before a decision, human or otherwise, can be made from it. None of this is malfunction. It is what a market of inventory-managing machines looks like from the outside.

The Loop That Closes on Itself

In a market of people, the chain from observation to action ran through a human whose attention was finite, who reacted at the speed of thought, and who did not always react the same way twice. That slowness and inconsistency were a form of damping. They limited how strongly and how quickly one participant's action could propagate into another's. Machines remove both. Each system observes the book, infers a state, decides, and acts; the action changes the book; other systems observe the change, infer, decide, and act in turn; and the first system then observes a market it helped to create. Figure 3 lays the loop out as a sequence of stages, with the rough time each takes.

Control engineers would describe such a loop by three properties: its gain, or how strongly a change in the book translates into new orders; its delay, or how long that translation takes; and the correlation of its elements, or how many systems respond to the same change in the same way. A loop with high gain, short delay, and correlated elements overshoots and oscillates; a loop with low gain, long delay, and independent elements settles. People supply the second kind of loop. Systems, unless they are deliberately designed otherwise, supply the first. The important point is that no individual system needs to be misbehaving for the loop as a whole to behave badly. Each can be prudent, well-tested, and compliant, and the population can still produce dynamics that none of its members would choose.^[2]

There is a subtler consequence for anyone who builds the systems. Every model in the loop was estimated on data that the other systems generated. The statistical regularities it learned, the shape of order-book imbalance before a price move, the typical decay of a temporary impact, the correlation between two instruments at short horizons, are partly properties of the underlying assets and partly artifacts of the rules the other machines were following at the time. When the population changes, the regularities change with it. Every model of the market is, in part, a model of the other models, and it is only as durable as they are.

What Emerges

Loops with those properties produce recognizable behaviors that nobody designed. Some have names because regulators have had to define them; most are simply what happens when many similar rules run against each other at speed. Figure 4 catalogs the ones we find most useful to be able to recognize, with the mechanism behind each and the signature by which it can be identified in the data. One entry, quote stuffing, is deliberate and prohibited in most jurisdictions; the rest are emergent, which is to say that they occur without anyone intending them and are correspondingly harder to legislate away.

Of these, correlated withdrawal deserves the most attention, because it is the mechanism by which a shock becomes a crash. Every market maker has a rule for when to stop: if realized volatility exceeds a threshold, if recent fills look informed, if inventory reaches a limit, widen the quote or pull it. Each rule is individually sensible; a system that never withdrew would be picked off. But if many providers share similar thresholds, a single event can trip them all in the same instant, and the systems that remain then see a thinner book, which raises their own estimates of toxicity and trips their rules too. The 2010 flash crash unfolded in minutes and the dislocation of March 2020 over weeks; they differ in speed and scale, but the underlying mechanism, prudent rules firing in concert, is the same in kind.^[3] Figure 5 simulates the difference between a population of liquidity providers with diverse thresholds and one whose thresholds are shared.

The recovery in that picture is as instructive as the withdrawal. Machines come back as quickly as they left, because their rules say to, and the quote after the event is often as tight as it was before. This is why the claim that automation has made liquidity worse is wrong on average and right in the tail. On an ordinary day the automated market is deeper, tighter, and cheaper than the market it replaced. In the moments that matter most, its liquidity can be gone before a person has finished reading the headline. The distribution of liquidity has not simply shifted; it has become bimodal, and the two modes are separated by the trip point of many similar rules.

How We Think About It

Three principles follow from taking this seriously. The first is to model the market as a population rather than as a counterparty. Our assumptions about depth, adverse selection, and impact are conditioned on regime: not only how deep the book is but how many independent decisions stand behind it, and how correlated their withdrawal rules are likely to be. Depth supplied by one kind of participant is worth less than the same depth supplied by several kinds, and we try to price the difference. The second is to treat our own systems as part of the loop. Our orders feed other systems' models, and our execution can be the event that trips their rules. We estimate our footprint, design execution to avoid resembling the signatures in Figure 4, and prefer to be a source of damping rather than gain. The third is to build for the tail: risk controls that assume displayed liquidity can vanish in milliseconds, circuit breakers that fire on the signature of a withdrawal rather than waiting for losses, and human judgment positioned where it is actually useful, which is at the level of parameters and regimes rather than individual orders.

The machines are not leaving, and we expect their share to keep rising as decision systems become more autonomous. The useful question is not whether machines trading with machines is good or bad for markets; it is what kind of dynamics a population of similar, fast, tightly coupled decision rules produces, and how to be one of the rules that does not amplify. A market of people was forgiving of individual error because its collective reactions were slow and diverse. A market of systems is far less forgiving of collective error, and the only defense is to understand the collective.