The Price Nobody Set

The Label

The opening exhibit is not a scandal. It is a laboratory result. Give reinforcement-learning agents a repeated pricing game, a common demand environment and nothing else — no channel to communicate, no instruction to cooperate, no code that mentions rivals at all — and they repeatedly learn to charge supra-competitive prices, sustained by punishment strategies: a deviation by one agent triggers a temporary price war, followed by a gradual return to the profitable equilibrium (Calvano, Calzolari, Denicolò and Pastorello, American Economic Review, 2020). The behaviour is learned, not programmed. There is no message to intercept, no meeting to prove, no employee who can be flipped, because the only thing the agents ever exchanged was the market itself: prices, posted publicly, where prices are supposed to be.

Precision about what this shows, held for the rest of the essay: controlled simulations demonstrate that coordination can emerge without communication; oligopoly theory has predicted for decades that transparency and repetition make it easier; early empirical work suggests the effect can appear in real markets, and is contested; everything about scale and prevalence remains hypothesis. The essay needs only the first two to make its argument. The others determine how urgent the argument is, not whether it is right.

When results like these surface in policy debate, the label arrives instantly: collusion, or in the trading context, manipulation. The legal label is familiar. The mechanism is not — and the distance between them is where this essay lives.

The Presumption of Authorship

The law of coordinated conduct can be read as a ladder — an analytical device, not a doctrinal hierarchy — and it is worth climbing slowly. The first rung is agreement: did firms communicate or reach an understanding? That is the classic cartel question, and the case law defines it broadly — a concerted practice is any form of coordination which knowingly substitutes practical cooperation for the risks of competition (Suiker Unie, 1975), and a single exchange of information can suffice (T-Mobile Netherlands, C-8/08, 2009). The second rung is intent: did an actor deliberately pursue the conduct? Some rules require it; others, importantly, do not. The third rung is the one this essay is about: attribution — can the legally relevant conduct be attached to an identifiable actor at all? Even a rule that cares nothing for intent still operates through an addressee, because a sanction has to land on someone. And the fourth rung is the phenomenon: emergence — an outcome arising from interaction without any participant having selected it.

The doctrine already knows where its ladder ends. Parallel conduct, by itself, has never established an agreement or concerted practice; the courts require that concertation be the only plausible explanation for it (Wood Pulp II, 1993). Conscious parallelism — each oligopolist watching the others and rationally declining to compete — is lawful, and the law made peace with that because sustaining parallelism across a market used to require exactly the conscious, communicated coordination the doctrine was built to detect. The difficulty changes when parallel conduct can be generated, stabilised and enforced without any of it.

So the precise claim, stated once and reused throughout: the difficulty arises where no individual firm has made the attributable decision the legal rule conventionally expects to identify, even though the interaction of individually deployed, individually supervised systems produces the regulated outcome. This is not a claim that nobody is responsible in the ordinary corporate sense. Firms chose their algorithms, set their objectives, supplied their data and continue to operate their systems. It is a claim about levels: the decision the rule wants to find does not exist at the level the rule addresses.

Three Mechanisms Hiding Under One Label

"Algorithmic collusion" is used for three different mechanisms, and the law's grip weakens as they are separated.

Coordination by instruction. Humans collude; algorithms execute. Two online sellers agree not to undercut each other and configure repricing software to enforce the pact — the UK's poster-seller cartel, prosecuted without attribution difficulty. The algorithm here is a fax machine with better throughput. An author exists, confessed, and was fined. The law is fully equipped, because the technology changed the speed of the conduct, not its structure.

Coordination through an intermediary. A common pricing vendor sits between competitors, ingesting their non-public data and returning aligned recommendations. This is the US rent-pricing litigation: the Department of Justice sued RealPage and major landlords over revenue-management software that pooled competitively sensitive information, and settled in late 2025 by consent decree — no fine, but conduct terms that read like a map of the theory: the company must stop using competitors' non-public data in its pricing recommendations and stop training its models on active lease data, under a compliance monitor (United States v. RealPage, proposed consent decree, November 2025). The author is displaced into the supplier, but findable: the hub-and-spoke logic recovers attribution because someone, somewhere, still aggregated private information. Note what the settlement's own architecture concedes — it regulates the data the algorithm may see and learn from, the substrate, because the conduct itself had no signable moment. One caution the debate keeps skipping: these proceedings involve pooled non-public information — conventional links. They are evidence for the second mechanism, not the third, and citing them as proof that fully independent systems already coordinate overstates what has been shown.

Coordination by interaction. No vendor, no pooled data, no instruction. Independently developed learning systems, each trained only on public market signals, each individually compliant, whose policies interact to generate the outcome the first two mechanisms were punished for. Enforcement practice has climbed the first two rungs. The third is where the addressee runs out — and where the rest of this essay takes place.

The Tape: Visibility Without Authorship

Coordination without communication has an infrastructure requirement, and oligopoly theory identified it long before machine learning existed: observability. Punishment strategies only work if deviation can be detected; detection requires seeing rivals' prices quickly and reliably; and the faster and cleaner the common signal, the cheaper the punishment and the more stable the coordinated outcome. The causal chain runs visibility → observability of deviation → ability to condition future action on rivals' behaviour → sustainable coordination. Nothing in that chain requires a meeting.

Which brings the argument to the newest piece of European market infrastructure. On 14 September 2026, EuroCTP began operating the EU's first consolidated tape for shares and ETFs: pre-trade and post-trade data from roughly 130 venues and reporting arrangements, merged into a single stream, under ESMA's direct supervision, free to retail investors and academics, with the formal five-year operational period running from the end of the transition on 30 September (ESMA authorisation, 27 July 2026; operations commenced 14 September 2026). The tape exists for excellent reasons this library has already argued — fragmentation made European price discovery a subscription product, and consolidation repairs that. But the same design decision has a second reading. What a consolidated tape changes, relative to fragmentation, is the availability, standardisation and timeliness of the common signal: one feed, which any learning system in the market can observe, condition on and train against. Regulation did not merely permit that observability; it prescribed it, built it and switched it on two months ago. Whether coordination comes to run on it is precisely the question the tape's design never had to ask.

The dual effect has to be stated honestly, and early. Transparency serves price discovery, investor protection and competition between venues — and simultaneously changes the economics of coordination between the algorithms watching it. Both are true; neither cancels the other. Which is why the policy question is never "transparency or not". It is what information, at what granularity, to whom, with what delay — and in what form automated systems can consume it — a question the tape's designers answered for investor-protection objectives; the second reading was not the question they were asked. The rulebook, here as elsewhere, is building the condition it will later have to answer for — the same structure this library just traced in deposit funding, where the calibrations lean on a behaviour that automation deletes; here, the infrastructure supplies a signal that automation exploits.

Every firm can be individually compliant while the interaction is the thing no one governs.

The Market Maker's Daily State

None of this is unsupervised, which is what makes it interesting. The algorithmic-trading rulebook is genuinely demanding: a firm deploying algorithms must have tested them, must monitor them in real time, must be able to kill them with one action, must maintain systems designed to prevent its trading from creating or contributing to a disorderly market or being used contrary to market abuse law, and must self-assess against those requirements annually (MiFID II, Art. 17; Commission Delegated Regulation (EU) 2017/589). The supervision is real. It is also, structurally, supervision of the member — each firm's system, tested in each firm's environment, against each firm's obligations.

The phenomenon of Section III's third mechanism is a property of the ensemble. It does not live inside any firm's system; it lives between them, in the interaction of policies that were each individually lawful, individually tested and individually kill-switched. Firm-level controls produce individually supervised systems, and the interaction between them remains outside the supervisory object — not through negligence, but because supervision, like sanction, attaches to an addressee, and interaction has no owner. The market maker's day already looks like this: a human formally responsible for a quote stream no attention span can cover, ratifying what the machine decides — the state this library's essay on judgment described as a destination, arrived early.

Where the Law Reaches, and Where It Stops

Now the honest survey, starting with a correction to the debate's usual framing. The problem with market abuse law is not that it presumes intent. Several of its manipulation limbs are effects-based on their face — transactions or orders that give false or misleading signals, or secure prices at an abnormal or artificial level, are caught without any intent element, subject to a legitimate-reasons defence, and the regulation explicitly lists algorithmic and high-frequency strategies among its examples (Regulation (EU) 596/2014, Art. 12(1)(a), (2)(c)). On paper, MAR can describe the third mechanism's output: a price at an artificial level. What is unsettled is to whom the description attaches. Even an effects-based rule ultimately operates through the conduct of an identifiable person or undertaking — and when no individual system, viewed in isolation, appears to produce the prohibited effect, which conduct by which person satisfies the substantive elements is precisely what the framework does not say. The undertaking exists; the act the rule needs is the missing piece.

Competition law reaches further than its caricature, and stops for the same reason. The concerted-practice doctrine has been stretched over platforms before: in Eturas, travel agents who received a system message capping discounts were presumed to have joined a concerted practice unless they publicly distanced themselves or systematically ignored it (C-74/14, 2016). The tempting move is to extend that logic to learning systems — if your algorithm foreseeably coordinates, you accepted the coordination. But Eturas turned on a communicated message with an identifiable sender and a knowable content. Whether doctrines built around communicated or knowingly received information can translate to behaviour that is learned rather than instructed is precisely the open question, and foreseeability is exactly what a firm deploying a self-training system can honestly deny. The policy literature has circled this gap for nearly a decade without closing it (OECD, Algorithms and Collusion, 2017, and successor work).

Put the evidential question directly, because it decides everything: what exactly must an authority prove? That the firm chose the algorithm? That it could have foreseen the outcome? That it should have known the system would converge? Or that it continued to operate the system after observing the outcome? Each proposition is different, and none is equivalent to proving an agreement. The first three all strain in the same direction — if foreseeability becomes a substitute for agreement, ordinary strategic interdependence converts into liability; if it stays irrelevant, a genuinely harmful pattern escapes the enforcement model entirely. The fourth is the one that can survive the emergence problem, because it asks nobody to have selected the learned policy — only to have watched it work and kept it running. Deployment becomes a continuing decision rather than a past one, and knowledge acquired after the fact becomes the attribution hook that design-time innocence cannot dissolve. And the doctrinal pattern for it already exists: in Eturas the Court presumed participation from awareness plus continued conduct, rebuttable by public distancing, reporting to the authorities, or systematic deviation from the coordinated behaviour (C-74/14, EU:C:2016:42; cf. the participation presumption of Anic, C-49/92 P, 1999). The mirror image for a learning system writes itself: continued operation of a system observed to coordinate, without adjustment and without report, as tacit approval — with systematic deviation, the firm that retrains its pricer to break the pattern, as the rebuttal. It is a narrow bridge between the rungs, but it is the only one that carries weight.

The empirical stakes are no longer purely theoretical, and no longer settled either. The best field evidence to date comes from German retail fuel: after stations adopted algorithmic pricing, margins rose materially — on the study's estimates, by roughly a tenth on average and by close to a third in local duopolies where both competitors adopted, with the increase emerging only after a delay consistent with learning (Assad, Clark, Ershov and Xu, Journal of Political Economy, 2024). One market, one study, contested interpretation — the essay claims no more. But the shape of the result is exactly the shape the simulations predict, and the study identifies no conventional communication behind it.

Why the Obvious Fixes Fail

Three fixes dominate the debate, and each fails mechanically rather than rhetorically.

Ban the collusive algorithm. There is no such object. The coordinating property is a learned policy, not an inspectable code feature; the same architecture that learns to undercut in one environment learns to punish deviation in another. A prohibition would have to name the feature it prohibits, and the only honest candidate — learning from rivals' observable behaviour — describes every pricing system worth deploying.

Ban the outcome. Treat sustained parallel pricing as an infringement regardless of mechanism. But parallel pricing is also what lawful interdependence produces in any concentrated market; an outcome test cannot distinguish emergent coordination from ordinary oligopoly without abandoning the settled position that conscious parallelism, by itself, is not an infringement — a reversal whose costs would land well beyond algorithmic markets.

Mandate a human in the loop. Formal approval does not create substantive control. A human co-signing ten thousand quotes an hour is not a decision-maker; they are an attribution device — a name the sanction can finally land on, manufactured for that purpose. It restores the addressee and changes nothing about the conduct. Each of the three fixes, in its own way, makes the same move: it manufactures an author instead of governing a condition.

Regulating the Condition

If authorship cannot be restored, the answer is not to abandon firm-level responsibility but to add a second layer to it. The first layer asks whether each participant complies with the rules addressed to it — that layer already exists, and it is genuinely demanding. The second asks whether the market's architecture creates predictable interaction effects that no participant can see alone — that layer does not exist. Four instruments for building it, in rising order of originality.

Detection. Suspicious-transaction reporting is firm-level by construction: each firm reports anomalies in its own flow. An ensemble property is invisible at that level by definition. Interaction-level surveillance — cross-firm, cross-venue pattern analysis, sitting where the consolidated data already sits — is the minimum condition for even knowing whether the third mechanism is operating. The tape that aggravates the problem is also the natural instrument for watching it.

Disclosure. Supervisors cannot reason about systemic interaction between systems whose objectives, constraints and training regimes they cannot see. Firm-level algorithm governance already generates this documentation internally; the step is supervisory visibility into it, calibrated to understanding interaction rather than auditing code.

Intermediary accountability. Where a common vendor genuinely structures the conduct, conventional attribution still works and should be used to its limit — the second mechanism treated as the first. The RealPage decree shows both the reach and the boundary: attribution recovered because private data was pooled, and the remedy aimed at the data the algorithm may see and train on. Where the only common input is the public tape, that logic runs out — which is exactly where the fourth instrument, market design, begins.

Market-design parameters. The uncomfortable one. If the timing, granularity, aggregation and access profile of strategically relevant information — and the form in which automated systems can consume it — shape the economics of coordination — and the theory says they do — then those parameters are policy dials, not technical footnotes. Delay for some users and uses, aggregation of the most conditioning-relevant fields, throttled access tiers: a design toolkit in which deliberately coarsened information is one member, not the headline. This is not an argument against the tape. It is an argument that the tape's parameters were set to answer one question — investor protection — and are also the answer to a second question nobody asked.

The distributional point, carried as hypothesis rather than verdict: the firms best able to learn from the common signal may capture part of the value the resulting coordination creates, while the legal framework struggles to identify the conduct that generated it. Who benefits from the gap remains the right question. The essay does not pretend to have measured the answer — only to have shown why, under current categories, the framework does not clearly assign the obligation to anyone.

The object of regulation shifts from the actor to the interaction.

The Price Without an Author

Markets were regulated as conversations between decision-makers. They are becoming weather — outcomes produced by interacting policies, each individually chosen, none of which selected the result. A law organised around identifiable decisions can search such a market for an author, find none, and acquit — and the market outcome remains exactly as it was.

The answer is not to invent culprits. Manufactured authors — the co-signing human, the strained foreseeability doctrine — satisfy the form of enforcement while leaving the mechanism untouched. The answer is to admit that some market outcomes emerge from interaction, and to regulate the conditions under which they emerge: watch the ensemble, see the objectives, hold the intermediaries that exist, and treat the market's information architecture as the policy instrument it has quietly become.

The problem, in the end, is not that machines have become impossible to regulate. It is that some economically consequential behaviour now arises at the level of interaction, while the law is still organised around the level of the actor. The price nobody set is still a price. And the difficulty is not that no one can ever be held responsible — the essay's own bridge says otherwise. It is that responsibility for an emergent outcome cannot be inferred from the fact that the outcome exists; it has to be constructed, out of continuing duties, observable conditions and named parameters. Someone will have to govern how such prices come about, precisely because no one can be convicted of setting them.