Competition Law and Machine Learning Models for Cartel Screening

1. Introduction

Machine learning (ML) models for cartel screening refer to computational systems that analyse market and transaction data to identify patterns that may indicate possible cartelisation or coordinated conduct.

Traditional cartel enforcement often depends upon:

• whistle-blowers;
• leniency applications;
• inspections;
• emails and documents;
• witness testimony;
• suspicious communications;
• economic analysis.

Machine learning can supplement these methods by screening large datasets for unusual patterns in:

• prices;
• bids;
• quantities;
• market shares;
• customer allocation;
• tender participation;
• geographic allocation;
• supply patterns.

However, an important legal principle is:

An ML model can identify a lead or risk indicator; it does not by itself prove the existence of a cartel.

A competition authority must still establish the legal elements required by the relevant competition law.

2. Meaning of Cartel Screening

Cartel screening is the process of identifying markets or firms that deserve closer competition-law investigation because available information suggests possible coordination.

Traditional approach

Data → human economic analysis → investigation

ML-assisted approach

Large dataset → ML model → suspicious pattern → human investigation → evidence → legal assessment

Machine learning therefore works primarily as an investigative screening tool.

3. What Is a Cartel?

A cartel generally involves competing businesses coordinating rather than competing independently.

Common forms include:

1. Price fixing

Competitors agree on prices.

2. Bid rigging

Competitors manipulate procurement or tender outcomes.

3. Market allocation

Competitors divide:

• customers;
• territories;
• products.

4. Output restriction

Competitors coordinate production or supply.

5. Information exchange

Competitors exchange strategically sensitive information in circumstances that facilitate coordination.

These forms of conduct are among the central concerns of competition law.

4. Why Machine Learning Is Useful for Cartel Screening

Cartels can generate large quantities of economic information.

For example:

• prices may move together;
• winning bidders may rotate;
• competitors may avoid bidding against one another;
• market shares may remain unusually stable;
• geographical allocation may appear systematic.

An ML system can process millions of observations much faster than manual review.

Example

Suppose 500 companies participate in 20,000 public tenders.

An ML system can examine:

• who bids;
• who wins;
• bid prices;
• bid gaps;
• timing;
• tender specifications;
• geographic patterns.

It can identify combinations that deserve investigation.

5. Machine Learning vs Traditional Screening

ML should therefore generally supplement rather than replace legal and economic analysis.

6. Types of ML Models

A. Supervised learning

The model is trained using historical examples labelled as:

• cartel;
• non-cartel.

It then attempts to classify new markets.

Examples

• Random forests
• Logistic regression
• Support-vector machines
• Gradient boosting

B. Unsupervised learning

The model looks for unusual patterns without being given cartel labels.

Examples include:

• clustering;
• anomaly detection;
• principal-component analysis.

This is useful where confirmed cartel datasets are limited.

C. Neural networks

Neural networks can analyse complex relationships between:

• price;
• quantity;
• timing;
• bidding;
• market structure.

However, they may be less transparent.

D. Anomaly-detection models

These identify observations that differ significantly from expected competitive behaviour.

For example:

Competitors normally submit bids with substantial variation.

But:

A particular group suddenly submits unusually similar bids.

The system can flag the market for investigation.

7. Important Cartel-Screening Indicators

ML models may examine indicators such as:

Price indicators

• unusually parallel prices;
• reduced price dispersion;
• repeated identical prices;
• sudden price increases.

Bid indicators

• bid rotation;
• unusual winning patterns;
• repeated second-place bidders;
• identical bid formatting.

Market-allocation indicators

• geographic patterns;
• customer-specific patterns;
• product allocation.

Capacity indicators

• unusual production restrictions;
• coordinated capacity changes.

Communication indicators

Where lawfully obtained and processed:

• suspicious communication patterns;
• meeting networks;
• common contacts.

8. Parallel Pricing Is Not Automatically a Cartel

This is one of the most important principles.

Suppose five firms increase prices simultaneously.

That does not automatically prove an agreement.

Parallel pricing can result from:

• common input-cost increases;
• demand changes;
• common market information;
• exchange rates;
• regulation;
• supply shocks.

Therefore:

ML should identify suspicious patterns, not convert correlation into proof of collusion.

Human and legal investigation remains necessary.

9. Economic Evidence and Legal Evidence

An ML model can produce:

"This market has a high probability of unusual coordination."

But competition law asks a different question:

"Has legally sufficient evidence established an infringement?"

The distinction is critical.

ML output

Investigative intelligence

Legal finding

Evidence-based determination under the applicable competition law

The two should never be treated as identical.

10. Case Law: United States v. Socony-Vacuum Oil Co.

Case

United States v. Socony-Vacuum Oil Co., 310 U.S. 150 (1940)

Principle

The Supreme Court treated price fixing among competitors as a serious violation of the Sherman Act.

Relevance to ML screening

The case illustrates why unusual pricing patterns can be important to investigators.

However, an ML model identifying parallel prices cannot itself establish the agreement that competition law requires.

11. Interstate Circuit v. United States

Case

Interstate Circuit, Inc. v. United States, 306 U.S. 208 (1939)

Principle

The Supreme Court considered coordinated conduct involving competitors and circumstances from which an agreement could be inferred.

ML relevance

The case demonstrates that cartel investigations can involve circumstantial evidence.

ML may help investigators identify patterns from which further evidence can be sought.

But the model's output remains only one part of the evidentiary picture.

12. American Tobacco Co. v. United States

Case

American Tobacco Co. v. United States, 328 U.S. 781 (1946)

Principle

The Supreme Court recognised that unlawful agreement can sometimes be established through circumstantial evidence and coordinated conduct.

ML relevance

This is important for automated screening because cartels may not always leave a simple document saying:

"We agree to fix prices."

ML can identify combinations of circumstantial indicators requiring further investigation.

13. Brooke Group v. Brown & Williamson

Case

Brooke Group Ltd. v. Brown & Williamson Tobacco Corp., 509 U.S. 209 (1993)

Principle

The Supreme Court established important principles concerning predatory pricing and the distinction between aggressive competition and unlawful exclusion.

Relevance

ML models must avoid treating every unusual pricing strategy as cartel behaviour.

A firm may legitimately:

• lower prices;
• respond aggressively to competitors;
• change prices rapidly.

Therefore, screening models must distinguish competitive behaviour from suspicious coordination.

14. Wood Pulp

Case

A. Ahlström Osakeyhtiö and Others v Commission, Joined Cases 89/85 and Others (Wood Pulp) (1988)

Principle

The European Court considered coordinated pricing conduct involving international producers.

Relevance

Wood Pulp demonstrates the importance of analysing economic evidence concerning parallel behaviour while distinguishing lawful independent conduct from coordinated conduct.

For ML systems, this is particularly important because global markets may produce highly correlated prices even without unlawful coordination.

15. Eturas

Case

Eturas UAB and Others, Case C-74/14 (2016)

Principle

The European Court of Justice considered an online booking platform where a common electronic system communicated a message capable of restricting discounts offered by participating travel agencies.

The Court considered circumstances in which knowledge of the communication and participation in the system could contribute to establishing concerted practice, subject to the applicable evidentiary requirements.

ML relevance

Eturas is particularly significant for the digital economy because it demonstrates that:

technology can become the mechanism through which competitors receive or implement coordination-related information.

ML screening can similarly analyse digital-platform data for unusual coordinated behaviour.

16. AC-Treuhand

Case

AC-Treuhand AG v Commission, Case C-194/14 P (2015)

Principle

The Court confirmed that an undertaking can potentially participate in a cartel even where it is not itself operating as a competitor in the affected market, depending on its contribution to the cartel.

ML relevance

Cartel-screening systems should therefore not focus exclusively on obvious competitors.

Potentially relevant actors can include:

• platforms;
• intermediaries;
• consultants;
• service providers.

The legal responsibility of each actor must nevertheless be established individually.

17. Cartes Bancaires

Case

Groupement des cartes bancaires (CB) v Commission, Case C-67/13 P (2014)

Principle

The Court clarified that conduct should not automatically be classified as a restriction of competition "by object" unless the necessary legal standard is satisfied.

ML relevance

This provides an important safeguard against over-classification.

An algorithm detecting suspicious conduct cannot simply label it:

"cartel"

without proper legal analysis.

18. Piau

Case

Piau v Commission, Case T-193/02 (2005)

This case concerned rules affecting professional football agents and competition-law analysis.

ML relevance

It demonstrates the broader importance of assessing:

• market structure;
• restrictive rules;
• competitive effects.

For ML screening, market context remains essential.

A statistical anomaly without market context can easily produce a false positive.

19. Machine Learning and Bid Rigging

Public procurement is particularly suitable for algorithmic screening.

Example indicators

Suppose companies A, B, C and D repeatedly participate in government tenders.

The model identifies:

• A wins Tender 1;
• B wins Tender 2;
• C wins Tender 3;
• D wins Tender 4.

Further analysis shows:

• losing bidders submit unusually high bids;
• firms rarely compete seriously against one another;
• winning firms rotate systematically.

This could constitute a screening signal.

It does not, by itself, prove bid rigging.

Investigators would need to seek additional evidence.

20. Machine Learning and Price-Fixing Screening

A model could examine:

Pit=f(Costs,Demand,Season,Competitors,Time)P_{it}=f(Costs, Demand, Season, Competitors, Time)

It could identify price movements that cannot easily be explained by observable market conditions.

For example:

Input costs remain stable → competitors simultaneously increase prices → price dispersion sharply falls.

This might justify further examination.

But alternative explanations must be considered.

21. Cartel Screening Through Anomaly Detection

An anomaly-detection model could identify:

Normal competitive pattern

Prices vary substantially.

Suspicious pattern

Prices suddenly become highly similar.

The model assigns a risk score.

For example:

Market A → low anomaly level
Market B → moderate anomaly level
Market C → unusually high anomaly level

These scores should be used for investigative prioritisation, not as legal guilt scores.

22. False Positives

False positives occur when the model identifies innocent behaviour as suspicious.

Examples:

• common input-cost shocks;
• common regulatory changes;
• seasonal demand;
• identical public information;
• commodity-price changes.

Example

Oil prices increase by 20%.

Ten airlines increase ticket prices.

An ML model may detect strong price correlation.

But the correlation could result from a common cost shock rather than a cartel.

Therefore:

Correlation ≠ coordination.

23. False Negatives

False negatives occur when a cartel exists but the model does not identify it.

This can happen when:

• firms deliberately vary prices;
• coordination is sophisticated;
• data is incomplete;
• cartel members use indirect mechanisms;
• market conditions change.

Therefore, ML should never be treated as a complete substitute for:

• leniency programmes;
• inspections;
• interviews;
• documentary evidence;
• economic analysis.

24. Explainability

A competition authority should ideally be able to explain why an ML system flagged a market.

Less explainable

"The neural network says there is a 93% cartel probability."

More useful

"The model identified unusually high bid similarity, systematic bid rotation and repeated geographic allocation patterns."

The second approach is more suitable for investigative decision-making.

25. Human Oversight

Human review should generally remain central.

Recommended structure

ML model

↓

Risk signal

↓

Economist review

↓

Competition-law analysis

↓

Evidence gathering

↓

Investigation

↓

Legal determination

This prevents the machine from becoming the de facto decision-maker.

26. Data Quality

The quality of an ML model depends heavily on its data.

Problems include:

• missing information;
• inaccurate prices;
• incomplete tender records;
• inconsistent product definitions;
• changing market conditions;
• measurement errors.

Principle

Bad data → bad screening → potentially bad enforcement decisions.

Competition authorities should therefore validate datasets before relying heavily on model outputs.

27. Training Data Bias

Historical cartel datasets can themselves be biased.

For example, confirmed cartels may disproportionately come from certain industries.

If a model is trained only on these examples, it may perform poorly in:

• digital markets;
• emerging technologies;
• service industries;
• developing economies.

The model may therefore require continuous validation.

28. Dynamic Markets

Machine-learning models can become obsolete.

Markets change because of:

• new technologies;
• new competitors;
• mergers;
• regulation;
• supply shocks;
• changing consumer preferences.

A model trained on 2015 procurement data may not accurately interpret 2026 digital-market behaviour.

Therefore, model drift must be monitored.

29. Privacy and Data Protection

Cartel screening may involve:

• emails;
• communications;
• employee information;
• transaction records.

Authorities must comply with applicable:

• privacy laws;
• data-protection requirements;
• procedural safeguards.

Competition enforcement cannot simply ignore privacy because ML is being used.

30. Due Process

Where ML contributes to an investigation, affected businesses should have appropriate procedural protections.

Important principles include:

• transparency where legally required;
• opportunity to challenge evidence;
• access to relevant evidence subject to confidentiality rules;
• independent human assessment;
• reasoned decisions.

A company should not be penalised solely because an opaque algorithm classified it as suspicious.

31. AI and Autonomous Cartels

An especially difficult future question is whether pricing algorithms could independently produce coordinated outcomes.

Consider:

Firm A AI

↕
Firm B AI

↕
Firm C AI

The systems continuously observe market conditions and adjust prices.

They may eventually learn that aggressive price competition reduces profits and independently converge on stable pricing.

This creates a difficult competition-law question.

Important distinction

Autonomous coordination is not automatically equivalent to a legally established cartel.

Competition law must determine whether the applicable legal requirements for an agreement, concerted practice, or other infringement have been satisfied.

32. Algorithmic Tacit Coordination

Tacit coordination occurs where firms independently adapt to one another without an express agreement.

Competition law traditionally distinguishes between:

• lawful conscious parallelism; and
• unlawful agreement or concerted practice.

ML may make tacit coordination easier by allowing firms to:

• monitor rivals instantly;
• react to price changes;
• predict competitor behaviour.

This may create a policy challenge even where conventional cartel evidence is absent.

33. Competition Law and Algorithmic Pricing

Algorithms can legitimately improve pricing.

For example:

• airlines optimise seat prices;
• hotels adjust room prices;
• retailers respond to demand;
• logistics companies adjust rates.

These are not automatically anti-competitive.

The concern arises when algorithms are used in a manner that satisfies the legal elements of coordinated or exclusionary conduct.

34. ML Screening Framework

A competition authority could use the following process:

Stage 1 — Data collection

Collect lawful market data.

Stage 2 — Data cleaning

Remove errors and inconsistencies.

Stage 3 — Feature construction

Create indicators such as:

• price variance;
• bid rotation;
• market shares;
• bid gaps;
• customer allocation.

Stage 4 — Model analysis

Apply:

• anomaly detection;
• clustering;
• classification;
• time-series models.

Stage 5 — Human review

Economists examine flagged patterns.

Stage 6 — Legal assessment

Determine whether the facts could satisfy the relevant legal provisions.

Stage 7 — Evidence collection

Seek:

• communications;
• contracts;
• internal documents;
• witness evidence.

Stage 8 — Enforcement decision

Only legally sufficient evidence should support an infringement decision.

35. Competition Authority Risk-Scoring

ML may help authorities prioritise resources.

For example:

These should be treated as investigative indicators, not legal scores of guilt.

36. Benefits of ML Cartel Screening

1. Scale

Millions of transactions can be analysed.

2. Speed

Potential anomalies can be identified quickly.

3. Consistency

The same screening methodology can be applied across markets.

4. Early detection

Authorities may identify suspicious markets before harm becomes extensive.

5. Resource allocation

Investigative resources can be directed toward higher-risk markets.

6. Pattern discovery

ML may detect relationships that are difficult to identify manually.

37. Risks of ML Cartel Screening

1. False positives

Legitimate competition may be flagged.

2. False negatives

Sophisticated cartels may escape detection.

3. Algorithmic bias

The model may over-identify certain industries.

4. Lack of explainability

Complex models may be difficult to interpret.

5. Data errors

Poor data can distort results.

6. Automation bias

Investigators may trust the model too much.

7. Due-process concerns

Businesses may not understand why they were targeted.

38. Important Case-Law Table

39. Key Legal Principles

For examination purposes, remember:

1. ML screening is an investigative tool, not a legal adjudicator.
2. Statistical correlation does not automatically establish cartelisation.
3. Parallel conduct can have legitimate economic explanations.
4. Circumstantial evidence can be legally important.
5. Digital platforms can facilitate coordinated conduct.
6. Human oversight is essential.
7. ML models must be trained and validated carefully.
8. False positives and false negatives must be recognised.
9. Due process applies to enforcement using algorithmic tools.
10. Data quality directly affects enforcement reliability.
11. Explainability improves the credibility of ML-assisted investigations.
12. Competition authorities should combine ML signals with documentary, testimonial and economic evidence.

40. Future of ML-Based Cartel Enforcement

Future competition authorities may combine:

• machine learning;
• natural-language processing;
• graph analysis;
• transaction analysis;
• bid analysis;
• anomaly detection;
• network analysis.

For example, a system could combine:

Price data

Tender data

Company relationships

Communication networks

Market structure

to identify markets requiring investigation.

The resulting system could become a form of continuous competition monitoring.

Nevertheless, continuous monitoring should remain subject to appropriate legal authority, privacy safeguards and procedural protections.

41. Conclusion

Machine learning models for cartel screening represent an important development in modern competition-law enforcement.

Their principal function is to identify patterns, anomalies and risk indicators that may warrant investigation. They can be particularly useful in:

• public procurement;
• price-fixing investigations;
• bid-rigging detection;
• digital markets;
• algorithmic pricing;
• large transaction datasets.

Cases such as Socony-Vacuum, Interstate Circuit, American Tobacco, Wood Pulp, Eturas, AC-Treuhand and Cartes Bancaires demonstrate why automated screening must be combined with proper legal and economic analysis.

The fundamental principle is:

Machine learning may tell an authority where to look, but it should not by itself determine what the law has been violated.

An effective framework therefore combines ML detection + economic analysis + human investigation + legally admissible evidence + due process. This approach allows technology to improve cartel detection without turning statistical probability into a substitute for proof of an actual competition-law infringement.