Collusion Screens Explained: How Economists Detect Cartels in Price and Bid Data

Published 2026-09-19. Last updated 2026-09-19. Sources are listed and linked at the end of this page.

A collusion screen is a statistical test applied to ordinary market data, prices, bids, quotes, spreads or market shares, that flags conduct fitting collusion better than it fits competition. The OECD's roundtable on ex officio cartel investigations gives the working definition: "A screen is a statistical test based on an econometric model and a theory of the alleged illegal behavior, designed to identify whether collusion, manipulation or any other type of cheating may exist in a particular market, who may be involved, and how long it may have lasted" [1]. What a screen returns is a ranked list of markets, firms or contracts worth examining, not evidence of an agreement. The same OECD paper states the limit without hedging: screens "are just the starting point of a cartel investigation, and rarely provide evidence sufficient to prove the existence of a cartel" [1]. A positive screen should change where investigative effort goes, not what anyone alleges.

Why screens exist

The economic stake is large and measurable. A review of the theoretical and empirical literature on collusion reports that "A recent comprehensive survey found the median increase in price attributable to collusion to be around 25%" [2]. Against that, the record of naive pattern matching is poor. In the 1970s the US Department of Justice ran an "identical bids" unit that investigated procurement auctions in which identical bids were submitted. The FTC Bureau of Economics working paper that introduced the variance screen records the outcome: "In six years, the unit failed to uncover a single conspiracy" [3].

Modern screening replaced the search for an obvious signature with a statistical one. A 2022 survey sorts the field into three objects: "A collusive marker is a pattern in the data more consistent with collusion than competition. A structural break is an abrupt change in the data-generating process that could be due to cartel birth, death, or disruption. An anomaly is a pattern in the data that is inexplicable or inconsistent with competition" [4]. Almost every screen in use is one of those three in a particular statistical dress.

Variance and price dispersion screens

The best documented family measures how much prices move rather than how high they are. A cartel must suppress the individual responses to cost and demand that generate ordinary price variation, so collusive prices are unusually rigid, and the rigidity ends abruptly when the cartel does.

The founding case study is a bid rigging conspiracy in frozen seafood sold to a US defence procurement centre, using frozen perch from 1987 through September 1989 with the price of fresh perch as a cost benchmark. Across the collapse of that conspiracy, "the mean price decreased by 16% while the standard deviation of price increased by 263%", and "its coefficient of variation increased by 332% from collusion to competition" [3]. The level fell modestly and the variance exploded. Practitioners normalise by the mean because the coefficient of variation is scale free, so contracts of very different sizes compare directly.

Competition authorities have since published usable thresholds. Calculations by the Swiss Competition Commission on a road construction cartel in the canton of Ticino found a coefficient of variation averaging 0.03 across bids during the cartel phase, with almost no rigged tenders above 0.05, rising to an average of 0.098 after the cartel broke down [5]. Those are concrete anchors for one sector in one country, and transplanting them without recalibration is a mistake.

The counterweight comes from the same paper. Having established the pattern on frozen perch, its authors searched for low variance pockets in retail gasoline around Louisville and reported a null: "We observe no such areas around Louisville in 1996-2002" [3]. A variance screen run over a real market will often find nothing, and that is the screen working correctly.

The false positive mode is easy to state. Anything that mechanically suppresses price variation mimics a cartel: regulated or indexed tariffs, long term contracts with fixed escalators, menu costs in retail, or open price leadership. None involves an agreement.

Structural break tests

A structural break test asks a narrower question: did the process generating these prices change abruptly, and when. Chow tests, multiple break estimators and regime switching models belong here, and they return a break date with a confidence interval, not a verdict.

Breaks are attractive because cartel birth and death are genuinely discontinuous. A worked example from procurement for generic pharmaceuticals shows the pattern at its cleanest: "Initially, the price was high and stable across tenders, and then suddenly, it was much lower and more variable" [4]. Both the level shift and the variance shift point the same way.

The problem is that breaks are common in competitive markets too. Entry by a large rival, the loss of a key input supplier, a currency move, a change in the procurement rulebook and a demand collapse all break the data generating process, and none is an agreement. The authors of that survey are explicit that this is tolerable only because of what a screen is for: "a screen is only asked to flag a market for further investigation" [4]. A break test must therefore be paired with a documented search for the cost or demand event that would explain it, dated against the estimated break.

Asymmetric price transmission, the rockets and feathers pattern

Asymmetric price transmission is the finding that output prices rise quickly when input costs rise and fall slowly when they fall. It was named for retail fuel, where a widely cited study summarises its own result as follows: the authors "test and confirm that retail gasoline prices respond more quickly to increases than to decreases in crude oil prices" [6].

It is frequently offered as a collusion marker, on the theory that firms coordinate quickly on the way up and hesitate on the way down for fear of triggering a price war. One finding cuts hard against it. A discussion paper from the US Department of Justice Antitrust Division's Economic Analysis Group notes that "Peltzman also found asymmetric pricing to be as common in unconcentrated industries as it was in concentrated industries" [7]. If the pattern appears just as readily where no one could plausibly coordinate, its value as a standalone marker is close to zero. Competitive explanations include inventory holding costs, menu costs, consumer search behaviour and production lags.

Asymmetry therefore belongs in a screening battery as a weak corroborating signal, never as a lead indicator, and a market showing nothing else should not be prioritised on it.

Parallelism and price change synchronisation

Parallel pricing is the pattern lawyers are asked about most and the one screens handle worst, because in many markets it is the competitive prediction. A former Acting Chairman of the FTC described a retail fuel market plainly: "there is complete price transparency because everybody can see the prices everyone else charges just by looking at those big signs" [8]. Under those conditions, matched prices are what a textbook expects of firms competing.

The same remarks locate the offence in the agreement rather than the pattern: "That glue is an express agreement among competitors, and it can overcome problems with insufficient price transparency, product differentiation, too many competitors and the like" [8]. A parallelism screen that finds identical prices has found the thing both hypotheses predict.

What carries more information is synchronisation in the timing and size of changes rather than in the level. If several firms change prices on the same day by the same amount, repeatedly, and the changes lead rather than follow any observable common cost, the coincidence becomes harder to explain. Screens here measure the distribution of intervals between rivals' price changes against what independent adjustment would produce. Common cost shocks, a published index everyone references, and open price leadership are the standing false positive modes.

Bid rigging screens for procurement

Procurement is where screening has the strongest track record, because tender data are structured and losing bids are often recorded. The OECD guidance sorts schemes into cover bidding, bid suppression, bid rotation and market allocation, and names the one screens meet most often: "Cover (also called complementary, courtesy, token, or symbolic) bidding is the most frequent way in which bid-rigging schemes are implemented" [9].

Each scheme leaves a different trace. Cover bidding distorts the distance between the winning bid and the losing bids, so screens measure the gap between the two lowest bids relative to the dispersion of the losing bids. Bid rotation shows up as a win sequence too regular for independent bidding. Market allocation shows up geographically, as firms that never meet in the same tender despite operating in the same region.

The classic bid rigging screen tests whether losing bids behave like real attempts to win. Studying state highway construction contracts on Long Island in the early 1980s, one study found "the rank distribution of higher cartel bids was unrelated to similar cost measures, and differed from the distribution of the low cartel bid" [10]. A genuine bidder's price tracks its own costs and capacity. A phantom bidder's tracks whatever number it was given.

Some flags need no statistics at all. In one set of 139 road paving tenders where the procurer set both a maximum and a minimum allowed bid, 123 winning bids fell between 91 and 95 percent of the maximum allowed, and the other 16 equalled the minimum, with nothing in between [4]. A bimodal distribution with an empty middle is not a competitive outcome.

The strongest published demonstration that screens can end in enforcement comes from Switzerland. Analysts took a road construction data set of 282 contracts let between 2004 and 2010, covering roughly 1,500 bids from 138 firms worth about CHF 216 million in winning bids, and applied combined variance and cover bidding tests with no prior information about collusion. Their design point was that real cartels are rarely comprehensive, so the method targets "collusion which does not involve all firms and/or all contracts in a specific data set" [5]. The Swiss Competition Commission opened an investigation in 2013 and sanctioned eight firms in 2016 [5].

Public tools have followed. The UK's Competition and Markets Authority published a free Screening for Cartels tool, of which it says: "The software uses algorithms to spot unusual bidder behaviour and pricing patterns which may indicate that bid-rigging has taken place" [11]. It also states: "In some cases, this kind of cartel can raise prices by as much as 30%" [11].

Method and limits

Screens generate hypotheses, not findings, and the gap between those two is where most misuse happens.

Start with what accuracy figures say. Applying penalised regression and ensemble methods to 584 Swiss tenders labelled as collusive or competitive, researchers found that "the algorithms correctly classify 84% of the tenders, as either collusive or competitive" [4]. That is a good result on a balanced, labelled sample. It says nothing about performance in the field, where the sample is not balanced.

Work the arithmetic. Suppose a screen with 84 percent sensitivity and 84 percent specificity runs across 1,000 tenders in a market where 2 percent are rigged. The 20 collusive tenders yield about 17 catches. The 980 clean tenders yield about 157 wrong flags. The flag list holds 174 items, of which 17 are real. Precision is about 10 percent, so roughly nine out of every ten flagged tenders are innocent. Nothing is wrong with the screen. The base rate did that, and no amount of model tuning removes it. Halving the false positive rate to 8 percent still leaves 78 false flags against 17 true ones.

This runs against the instinct to treat a flag as an accusation. A screen output is a work allocation instrument, and the correct response to a flag is to look for the innocent explanation first, because on the prior probabilities that is what you will usually find.

Transparency about error rates is not standard practice, including among authorities. A legal analysis of the CMA tool notes that nothing was published "concerning the precision, recall and F1 score of the different algorithms in the SfC tool" [12], which makes independent assessment of its field performance impossible. Any screening exercise should report its thresholds, its assumed base rate, its expected precision and the markets it was calibrated on.

Three further limits belong in any engagement letter. Screens cannot distinguish explicit from tacit coordination, which the OECD identifies as a specific source of false positives, since tacit parallel conduct without an agreement is generally lawful [1]. Screens are evadable: a cartel that scales its bids proportionally can satisfy independence tests while still rigging outcomes [2]. And an inconclusive result is often a statement about the data rather than about the market.

The OECD's summary is the right note to close on: "Cartel screens can produce false positives (flagging cases which do not merit further scrutiny) or false negatives (failing to identify collusion in a particular market)" [1]. Both errors are permanent. A screening programme is a way of spending scarce investigative attention better, and it should be judged on whether it improves that allocation, not on whether it is ever wrong.

Frequently asked questions

Is conscious parallelism illegal?

Not by itself in most systems. Matching a rival's price after observing it is the expected outcome in a transparent market for a homogeneous product. What competition law reaches is the agreement that holds coordination together, described in FTC remarks as the "glue" without which firms would often be unable to sustain parallel prices [8]. A parallelism screen therefore identifies conduct consistent with both lawful and unlawful explanations.

How do you prove bid rigging?

Not with a screen. Proof comes from agreement evidence, which in practice means documents, communications records, leniency applications or witness testimony. The screen's role is to justify the step that obtains that evidence. As the OECD roundtable records, screening results are commonly treated as sufficient to authorise an inspection, while rarely amounting to proof [1].

What is the difference between price fixing and bid rigging?

Price fixing is agreement on the prices charged in ordinary sales. Bid rigging is the same conduct inside a tender process, where the agreement determines who wins and what the losers submit. The distinction matters for screening because tenders record losing bids, which gives a screen far more structure to work with than a list of transaction prices.

How accurate are cartel screens?

On labelled research samples, reported correct classification rates reach 84 percent for tender level machine learning screens [4]. Field accuracy is different and generally unpublished: a low base rate pushes precision down sharply even when sensitivity and specificity are high. Ask any vendor or authority for precision at a stated base rate, not for accuracy.

Does cartel screening with machine learning work?

It improves feature selection rather than changing what a screen is. Supervised models trained on tenders known to be collusive or competitive extract the most informative markers automatically, and reported results on Swiss data are strong [4]. The requirement is labelled training data from resolved cases, and models calibrated on one country, sector and auction format should not be assumed to transfer.

What is a variance screen for collusion?

It is a test for prices that move too little. It compares the dispersion of prices or bids, usually the coefficient of variation, against a competitive benchmark or across a suspected cartel break. In the founding study the standard deviation of price rose by 263 percent when a bid rigging conspiracy collapsed, while the mean fell by only 16 percent [3].

Sources

1. Ex officio cartel investigations and the use of screens to detect cartels, OECD Competition Committee, DAF/COMP(2013)27, 2013. https://www.fne.gob.cl/wp-content/uploads/2014/07/2013-Ex-officio-cartels-investigation-3569-KB1.pdf
2. Detecting Cartels, Johns Hopkins University working paper, December 2004, revised July 2005. https://danielmorochoruiz.wordpress.com/wp-content/uploads/2018/02/harrington_detecting-cartels_wp526harrington.pdf
3. A Variance Screen for Collusion, FTC Bureau of Economics Working Paper No. 275, March 2005. https://www.ftc.gov/sites/default/files/documents/reports/variance-screen-collusion/wp275_0.pdf
4. Cartel Screening and Machine Learning, Stanford Computational Antitrust, 2022. https://law.stanford.edu/wp-content/uploads/2022/08/harrington-imhof-2022.pdf
5. Screening for Bid-rigging. Does it Work?, CRESE Working Paper No. 2017-9. https://crese.univ-fcomte.fr/uploads/wp/WP-2017-09.pdf
6. Do Gasoline Prices Respond Asymmetrically to Crude Oil Price Changes, Quarterly Journal of Economics, February 1997. https://cameron.econ.ucdavis.edu/research/qje97.html
7. Why Prices Rise Faster than they Fall, US Department of Justice Antitrust Division, Economic Analysis Group Discussion Paper EAG 09-4, July 2009. https://www.justice.gov/sites/default/files/atr/legacy/2009/07/28/248396.pdf
8. Should We Fear The Things That Go Beep In the Night? Some Initial Thoughts on the Intersection of Antitrust Law and Algorithmic Pricing, US Federal Trade Commission, 23 May 2017. https://www.ftc.gov/system/files/documents/public_statements/1220893/ohlhausen_-_concurrences_5-23-17.pdf
9. Guidelines for Fighting Bid Rigging in Public Procurement, OECD. https://uohs.gov.cz/download/Sekce_VZ/Metodiky/Bid_rigging_guidelines_OECD_en.pdf
10. Detection of Bid Rigging in Procurement Auctions, NBER Working Paper 4013, March 1992. https://www.nber.org/papers/w4013
11. CMA launches digital tool to fight bid-rigging, UK Competition and Markets Authority. https://www.gov.uk/government/news/cma-launches-digital-tool-to-fight-bid-rigging
12. 'Screening for Cartels' in Public Procurement: Cheating at Solitaire to Sell Fool's Gold?, University of Bristol Law Research Paper Series, Paper #006 2019. https://www.bristol.ac.uk/media-library/sites/law/research/'Screening%20for%20Cartels'%20in%20Public%20Procurement_Cheating%20at%20Solitaire%20to%20Sell%20Fool's%20Gold.pdf

Read this page as Markdown