Skip to content
Crypto.world

Research

Collusion with Competitive Marginals: Price-Level Audits Are Blind by Construction

Empirical work on algorithmic collusion asks one question of the data: are prices supracompetitive? The authors construct a conspiracy that is profitable yet answers "no" to that question, because it…

Collusion with Competitive Marginals: Price-Level Audits Are Blind by Construction
arXiv
2607.26385
Published
2026-07-29
Authors
Xin Xu, Chengrui Wu, Jiayu Lu, Kaizhen Tan, Siru Tao, Hanzhe Hong

AI summary

Overview

Research area: Algorithmic collusion and market design, sitting at the intersection of game theory, theoretical econometrics, and empirical crypto-market measurement (Ethereum MEV-Boost block-building auctions).

Technical level: Advanced. The core results are short proofs built on copula theory and the supermodular order, and the empirical work uses cross-fitted residual regressions, clustering, permutation nulls and family-wise error control. The prose is unusually accessible, but the argument depends on careful distinctions between marginal laws, process laws and joint laws.

Scope: A single paper arguing that all price-level collusion screens are powerless by construction against a profitable conduct that leaves every bidder's marginal bid distribution unchanged, with supporting evidence from simulation, 60 language-model agents and 24 days of Ethereum relay data.

What This Paper Is About

Empirical work on algorithmic collusion asks one question of the data: are prices supracompetitive? The authors construct a conspiracy that is profitable yet answers "no" to that question, because it never moves any single agent's price distribution—it changes only the dependence (the copula) between agents' unexplained bid components, and the agents need not communicate at all. The paper's goal is to show that this is not a low-power problem that more data can fix, to demonstrate the mechanism in real language-model agents and a live auction market, and to redirect regulation from detecting conduct toward counting how many independent decision-makers a market actually contains.

Key Contributions

  1. A zero-communication construction with exactly competitive marginals. Using a pseudorandom function keyed on public round data, coalition members generate a shared draw each round; a single scalar ρ tunes the joint law from independence (ρ = 0) to comonotonicity (ρ = 1) while every member's own marginal law stays exactly at F_i. The construction is profitable because seller revenue is the maximum bid, and the comonotone coupling minimises the expected maximum—in the paper's illustrative two-bidder case, expected revenue falls from 0.59σ to 0.28σ, a 52% drop.

  2. An exact impossibility result for marginal audits. Proposition 1 shows that any test reading a single bidder's history has power exactly equal to its size α, for every coupling strength up to comonotonicity, verified in simulation with 200 Monte Carlo trials per cell across coupling strengths and sample sizes up to 2 × 10⁴. A Bonferroni battery of six distributional statistics reaches power 0.020 at ρ = 1.0 and N = 2 × 10⁴ against nominal α = 0.05, while the pairwise test on the same data reaches power 1.

  3. Replication in real language-model agents. Twenty models from nineteen independent developers, three deployment prompts each (sixty agents), 200 rounds at each of five sampling temperatures, show that two deployments of one model couple at +0.053 under the strongest auditor (clustered 95% interval [0.030, 0.078]) while the 1710 cross-model pairs sit at +0.0001—a contrast that survives an auditor seeing every order feature and fitted out of sample.

  4. A market calibration plus a positive regulatory proposal. On 24 days of Ethereum MEV-Boost relay data (77,684 bids, 5,408 slots, 39 builders), the honest population of bidder pairs is itself so dependent that a screen held at a 5% false-positive rate must sit above +0.50 to +0.81, which is 20 to 32 times the family-wise sampling threshold and does not fall as the audit window grows. Since conspiracy and lawful multi-identity operation are behaviourally indistinguishable there, the authors propose counting entities: resolving 40 bidding identities into 23 operators raises the Herfindahl index by +247.5%, and adding behavioural clusters reaches +324.5%.

Main Findings

  • Marginal audits are blind by identity, not by low power. Because the shared draw is refreshed independently each round, the entire single-bidder process law—not merely the per-round marginal—is unchanged. Power equals size for every ρ up to 1, and no sample size repairs it. Simulation confirms a flat line at α = 0.05 from ρ = 0 to ρ = 1 at N = 200, 1k, 5k and 20k, while pairwise tests reach power 1 with resolvable coupling improving like N^(−1/2).

  • The blindness is bought exactly by the shared draw being non-persistent. Make the shared signal an AR(1) process instead, and the bidder's own series inherits the autocorrelation: at persistence 0.5 its lag-one autocorrelation is 0.40, and at 0.9 even a marginal goodness-of-fit test rejects. Serial dependence in one agent's residuals is therefore evidence of a different, detectable object.

  • Antithetic turn-taking is the wrong intuition. For uniform marginals and two bidders, the expected maximum is 1/2 under comonotonicity, 2/3 under independence and 3/4 under antithetic coupling. What costs a coalition money is the dispersion of its joint bids, so bidding alike beats taking turns.

  • The effect is comparable in size to what the price-level literature reports. That literature reports lifts of order 22%, falling to 7 to 10% under heterogeneity. On real relay data, a five-member coalition takes 4.6% of seller revenue at the lower end of the cleaned ambient floor and 10.6% at the uncleaned one—the same order of magnitude, and invisible to the standard tests.

  • Regret-based audits fail too, but for a different reason. Under a cross-fitted estimator, the competitive conduct at the same marginals already carries calibrated regret of order σ², and the conspiracy's regret R(ρ) is nonpositive in excess of that and decreasing in ρ. The coupling hides inside the market's own irreducible dispersion. This claim is conditional: it holds when that dispersion is off-best-response noise, which the measured κ makes plausible without proving.

  • Coupling is base-specific and weakens with sampling temperature. At temperature 0.3, same-model pairs correlate at +0.110 (auditor A), +0.059 (B) and +0.053 (C); cross-model pairs at +0.023, +0.000 and +0.0001. Correlation falls from +0.147 to +0.028 under (A) and from +0.070 to −0.004 under (C) as bid dispersion rises from 10.1 to 18.1, with permutation tests on temperature labels giving p = 5 × 10⁻⁵ for (A), p = 0.003 for (B) and p = 0.002 for (C). Developer-level means at T = 0.3 range from +0.17 down to −0.05 under the strictest auditor, so the mechanism is real on average rather than universal.

  • Forecasting and bidding differ, which sharpens the mechanism. Prior work measures cross-family forecasting correlation of 0.77 to 0.78, where a shared ground truth drives agreement; in bidding, cross-model correlation is indistinguishable from zero under a fully specified auditor because no ground truth exists and idiosyncratic strategy dominates.

  • The market's ambient dependence floor does not shrink with more data. Observed 95th-percentile pairwise residual correlation is +0.829 at a 100-round overlap requirement and +0.909 at 1000, while the permutation null falls from +0.130 to +0.050, and to +0.028 at 3000. Holding the population fixed settles composition effects: among the 55 pairs sharing at least 3000 slots, subsampling rounds over a thirtyfold range moves the null from +0.153 to +0.028 while the observed floor moves only from +0.910 to +0.893. On this market κ = 0.895, so 44.5% of cross-bidder dispersion is left unexplained by public covariates.

  • Three repairs fail. Entity resolution moves the floor only from +0.916 to +0.905, even though one operator runs six identities and same-operator pairs sit at median dependence +0.906. Removing five principal components carrying 85.8% of variance moves the observed 95th percentile to +0.825 while inflating the null from +0.114 to +0.233—losing power rather than gaining validity. More data is ruled out because the threshold contains no N.

  • The signature already exists lawfully in the data. Five identity pairs bid at correlation up to +0.992 while almost never taking top-of-book from each other, at leapfrog rates of only 0.028 to 0.060. Each such pair belongs to a single operator, and a screen calibrated to catch the conspiracy would flag these first.

  • Counting works; conduct detection does not. Pairwise bid-residual correlation recovers label-identified same-operator pairs at AUC 0.951, with an operator-clustered bootstrap interval of [0.754, 0.993] and leave-one-operator-out range 0.926 to 0.984. At threshold 0.80, recall is 95.8% at an apparent false-positive rate of 5.4%; permuting operator labels returns AUC 0.504, restricting to labelled pairs gives 0.957, and overlap alone gives 0.434.

  • Concentration estimates depend heavily on the counting rule. The Herfindahl index of bid leadership is 0.0655 over 40 bidding identities, 0.2277 over 23 entities (+247.5%), and 0.2782 over 20 units when behavioural clusters are added (+324.5%). Behavioural clustering alone is the weakest method, with a lift spanning +42% to +143% and +63% to +189% over tie-break order alone, so it carries no point estimate. Against size-matched random groupings the lift remains real (p = 0.0008).

  • Detection has costs on both sides of the dichotomy. A coalition facing a submission-timing screen can stay temporally separated, which screens at a floor of +0.499 rather than +0.905 and removes about half its rent, or co-submit and present as one operator's identity cluster, forfeiting its claim to independent competition. Covert rent as a percentage of proposer revenue runs 0.92%, 4.34% and 10.56% for coalitions of 2, 3 and 5 at the +0.905 floor, versus 0.46%, 1.81% and 4.57% at +0.499, on a panel with 103.2 ETH per day of proposer revenue.

Methodology in Plain English

The argument is built in three layers.

Layer one is a modelling choice. Each bid is split into a part an auditor can reconstruct from public data (round effects, persistent bidder effects, loadings on public covariates) and an unexplainable residual. The authors then define a "conduct profile" as a joint law for those residuals whose marginals are fixed at their competitive values. A conspiracy is simply a joint law under which the residuals of coalition members are dependent. Because the marginals never move, any conspiracy must live entirely in the dependence structure—the copula—and nothing else.

Layer two is a construction. To show such a conspiracy needs no communication, the authors have coalition members compute a shared random draw each round from a pseudorandom function keyed on public round data. Each member then transforms that shared draw plus its own private noise through its own marginal distribution. A single parameter ρ controls how much weight the shared draw gets, and the transformation is chosen so that the resulting marginal is exactly F_i for every ρ. Then they prove the key point: because the shared draws are independent across rounds, one bidder's own time series is statistically identical to what it would be under competition. Any test reading one bidder therefore rejects at exactly its false-positive rate. Two further propositions handle the economics (comonotone coupling minimises the seller's expected revenue, by a classical supermodular-order argument) and the regret-based audits (a cross-fitted estimator shows the conspiracy does not raise regret above what the honest, noisy market already carries). Simulations with 200 trials per cell verify the power claims across coupling strengths and sample sizes.

Layer three is measurement. For the language-model experiment, twenty models from nineteen developers are deployed under three prompts each across five temperatures, bidding in a first-price auction over synthetic orders. Bids are regressed on order features using three auditors of increasing flexibility, with the strictest using all six features plus squares and all pairwise interactions (27 terms) fitted by five-fold cross-fitting, and confidence intervals clustered by developer. For the market calibration, the authors use Ethereum's MEV-Boost block-building auction—the only free, public, live market they know of where both sides of every quote carry a persistent identity. Builders are bidders, proposers are sellers, and the relay-published bid value is the bid. They compare observed pairwise residual correlation against a permutation null, hold the set of pairs fixed while subsampling rounds to separate sampling error from composition, clean the population of same-operator pairs using the union of two independent labellings while deliberately excluding correlation-based clustering to avoid circularity, and then measure concentration under alternative counting rules.

Notably, the paper reports checks written specifically to attack its own earlier results, and the code release contains every script, result file and generated bid.

Why This Matters

Impact on research. The paper reframes a literature that has measured one thing—price levels—since Calvano et al. (2020). Its central claim is not that existing methods are underpowered but that they are structurally unable to see a particular profitable conduct, which changes what a negative empirical finding means. It also draws a clean line between price-level audits and the regret-based audits of Hartline, Long and Zhang (2024) and Hartline, Wang and Zhang (2025), showing the two fail for different reasons rather than treating "auditing" as one thing. The result that the detection floor is set by the spread of the honest population rather than by sample size is a general econometric point that applies well beyond collusion screening.

Real-world applications:

  • Procurement and auction design. Where bidders use common model supply or shared infrastructure, a regulator checking only whether bids exceed a competitive benchmark will conclude no collusion even when the seller's revenue has fallen. The paper's framing suggests auditing the dependence structure of residuals, or at minimum recognising the limitation.
  • Model supply-chain governance. Sampling temperature is a deployment parameter a regulator could mandate; the finding that coupling falls monotonically as temperature rises (p = 0.002 under the strictest auditor) turns it into a candidate mitigation lever, though the authors are careful to note the rent scale is not comparable across auditors.
  • Market concentration reporting. Entity resolution and behavioural clustering from public bid streams are both feasible and materially change measured concentration (+247.5% and +324.5% respectively), which matters anywhere market-share statistics are computed from raw identities.
  • Crypto market surveillance. For MEV-Boost specifically, the finding that five identity pairs bid at correlation up to +0.992 with leapfrog rates of 0.028 to 0.060 shows that the conspiracy's statistical signature is already present as ordinary lawful operator behaviour, meaning surveillance thresholds must be calibrated against that ambient floor rather than against independence.

Industry relevance. Any firm deploying foundation-model agents into pricing, bidding or procurement decisions inherits an upstream coupling channel it may not control and may not know about—the rent can accrue to a bidder-adjacent third party while nominal competitors bid in good faith. For crypto infrastructure, the paper quantifies how much measured market concentration depends on the counting rule, which directly affects how exchanges, relays and researchers report competitive structure. The "dual use" section is explicit: the mechanism arises by default from shared model supply rather than deliberate engineering, so the paper describes a condition markets are already in rather than a new capability.

Future Directions

  • Test temperature as a mandated dial at scale. The authors show the coupling weakens with temperature across model families but note the finding is not at conventional significance with five temperature points in every specification, and explicitly call for testing at scale since temperature is a parameter a regulator could set.

  • Break the entity-labelling ceiling. The labelling that grounds the detector and the concentration figures joins to builders only through winning blocks, so a builder that never won cannot be labelled and no amount of waiting fixes this; pooling to a longer panel makes coverage worse. A third, independent source of operator identity would let every identity-conditioned claim be validated.

  • Derive thresholds instead of sweeping them. The timing gap, clustering thresholds and screen operating points come from monotone sweeps reported as trade-off curves rather than derived from a loss function.

  • Prove or refute the regret claim in general. Part (iii) of Proposition 4—that conspiracy regret does not exceed the competitive conduct's regret and falls with ρ—is verified by a cross-fitted estimator across coalition sizes, outsiders, noise scales and value laws but left as a conjecture in general. Related open questions include whether a persistent common signal can be detected reliably, since the impossibility rests exactly on the shared draw being refreshed each round.

Target Audience

Competition economists and regulators working on algorithmic pricing and auction integrity; game theorists and econometricians interested in identification and detection limits; researchers building or auditing multi-agent LLM systems deployed in market settings; and crypto-market analysts studying MEV-Boost, builder concentration and entity resolution. The paper assumes comfort with copulas, supermodular order and cross-fitting, so the theoretical sections suit readers with graduate-level econometrics or decision theory, while the empirical and policy sections are readable by practitioners willing to take the propositions on faith. Readers looking for a practical detection tool will not find one; readers looking for a precise statement of what detection can and cannot achieve will.

Authors’ abstract

Empirical work on algorithmic collusion asks one question of the data: are prices supracompetitive? We show this can be answered "no" by a conspiracy that is nonetheless profitable. Consider bidding agents that couple only through the joint distribution of their unexplained bid components, leaving every agent's own bid law exactly at the competitive law. Any test whose input is a single agent's price or bid history then has power exactly equal to its false-positive rate, for…

The full abstract on arXiv
Read the original paper