Agent-Based Modeling for Antitrust Prediction Systems

1. Introduction

Agent-Based Modeling (ABM) is a computational method in which a market is represented as a collection of individual agents—such as firms, consumers, platforms, distributors, regulators and suppliers—whose decisions interact over time.

In competition law, an antitrust prediction system could use ABM to simulate questions such as:

whether firms are likely to coordinate;

whether an algorithm may facilitate collusion;

whether a merger could reduce competition;

whether a dominant platform could foreclose rivals;

whether a pricing strategy could produce exclusionary effects;

how competitors might react to a new entrant;

how a market could evolve after a regulatory intervention.

However, an ABM prediction is not itself proof of an infringement. EU competition law continues to require the applicable legal elements under Articles 101 and 102 TFEU. Article 101 addresses agreements, decisions and concerted practices restricting competition, while Article 102 addresses abuse of a dominant position. (Competition Policy)

The most important modern issue is therefore:

How can computational predictions be used as economic evidence without allowing a simulation to replace legal proof?

2. Meaning of Agent-Based Modeling

Traditional economic models often assume that market participants behave according to simplified mathematical rules.

ABM instead creates individual artificial actors.

For example:

Agent 1 — Firm A

chooses price;

observes competitors;

changes production;

learns from previous results.

Agent 2 — Firm B

changes price;

responds to Firm A;

attempts to maximise profit.

Agent 3 — Consumers

choose among products;

switch suppliers;

respond to price changes.

Agent 4 — Regulator

observes market outcomes;

changes enforcement parameters.

The model then allows these agents to interact repeatedly.

genui{"learning_viz":{"type_id":"MONOPOLY_INEFFICIENCY","locale_override":"en-US"}}

The visualization above illustrates the basic economic intuition behind monopoly harm; an ABM can go further by modelling how individual firms and consumers dynamically arrive at market outcomes.

3. ABM in Antitrust

An antitrust ABM might contain:

Firms

↓

Pricing decisions

↓

Competitor reactions

↓

Market demand

↓

Consumer decisions

↓

Market shares

↓

Repeated interaction

↓

Potential coordination/foreclosure

↓

Predicted competitive outcome

The model can run thousands of simulated market scenarios.

For example:

“What happens if Firm A acquires Firm B?”

or:

“What happens if three competing algorithms repeatedly observe each other's prices?”

4. Why Competition Authorities Might Use ABM

ABM can potentially help authorities analyse dynamic markets.

Traditional static analysis may ask:

What is the market share today?

ABM can ask:

What happens to the market over the next five years under different competitive conditions?

Possible applications include:

merger simulation;

cartel-risk assessment;

algorithmic pricing;

platform competition;

network effects;

entry and exit;

innovation;

foreclosure;

switching costs;

consumer behaviour.

The Commission's Article 102 framework already treats market definition, dominance, entry barriers, buyer power and vertical integration as relevant to assessing dominance. (Competition Policy)

ABM can potentially model several of those variables simultaneously.

5. ABM and Article 101 TFEU

Article 101 prohibits agreements, decisions by associations of undertakings and concerted practices that restrict competition. Examples include price fixing, market sharing and limitations on production or technical development. (Competition Policy)

ABM can be useful in examining a hypothetical question:

Can repeated algorithmic interaction produce coordinated outcomes even without an explicit written agreement?

For example:

Firm A's algorithm observes Firm B;

Firm B's algorithm observes Firm A;

both react rapidly to deviations;

price reductions are punished immediately;

prices converge.

The simulation may demonstrate that the market has a tendency toward coordinated outcomes.

But that does not by itself establish an Article 101 infringement.

Evidence concerning actual communications, instructions, awareness, conduct and market circumstances remains important.

6. Case Law 1 — Eturas

Eturas UAB and Others v Lietuvos Respublikos konkurencijos taryba

C-74/14, Court of Justice, 21 January 2016

This is one of the most relevant cases for computational competition analysis.

Several travel agencies used a common electronic booking system.

The system administrator sent a message concerning a restriction on online discounts, and the system automatically implemented the restriction.

The Court considered:

concerted practice;

automated systems;

communication through a computerised platform;

knowledge of the system's operation;

evidence and presumptions.

The case is particularly valuable for modern algorithmic antitrust because conduct can be facilitated through a common digital system, rather than a traditional face-to-face cartel meeting. (Infocuria)

ABM significance

An ABM could model:

platform message → algorithmic response → competitor behaviour → market outcome.

But the model cannot substitute for proof that the undertakings actually participated in the relevant concerted practice.

7. Case Law 2 — T-Mobile Netherlands

T-Mobile Netherlands and Others

C-8/08, Court of Justice, 4 June 2009

The case concerned information exchange between competitors.

The Court examined whether a meeting involving exchange of strategically sensitive information could constitute a concerted practice.

ABM relevance

ABM can model the economic impact of information exchange.

For example:

No information sharing

→ greater uncertainty

→ stronger price competition.

Compared with:

Strategic information sharing

→ reduced uncertainty

→ easier coordination.

The simulation can therefore help explain why information exchange might affect competitive conditions.

But again, the legal finding requires evidence concerning the actual conduct.

8. Case Law 3 — AC-Treuhand

AC-Treuhand v Commission

C-194/14 P, Court of Justice, 22 October 2015

AC-Treuhand concerned the role of a consultancy in cartel activity.

The Court confirmed that an undertaking can fall within Article 101 where its conduct contributes to implementing a cartel, even if it is not itself operating at the same level of the market as the cartel participants.

ABM relevance

This is important where computational systems are supplied by:

software companies;

algorithm vendors;

consultants;

data providers;

platform operators.

Suppose:

Algorithm provider

↓

supplies coordination-facilitating software

↓

Competing firms

↓

coordinated market behaviour.

An ABM could help investigate how the system affects market outcomes, but legal liability still depends on the applicable facts and Article 101 requirements.

9. Case Law 4 — Cartes Bancaires

Groupement des cartes bancaires v Commission

C-67/13 P, Court of Justice, 11 September 2014

The Court stressed the importance of distinguishing restrictions by object from restrictions requiring an effects analysis.

Why this matters for prediction systems

ABM is primarily useful where authorities need to investigate effects.

A model might examine:

prices;

output;

market shares;

entry;

innovation;

consumer welfare.

But an authority cannot simply say:

“The model predicts harmful effects, therefore Article 101 is automatically infringed.”

The correct legal classification still matters.

10. Case Law 5 — Intel

Intel Corp v Commission

C-413/14 P, Court of Justice, 6 September 2017

Intel concerned rebates granted by a dominant undertaking.

The Court held that where the Commission assesses whether a rebate system is capable of restricting competition, factors such as:

dominant firm's position;

market coverage;

conditions of the rebates;

duration;

amount;

possible foreclosure strategy;

equally efficient competitor analysis

may be relevant.

ABM significance

An ABM can simulate:

“What happens to competitors if the dominant firm introduces this rebate?”

The model could estimate:

rival exit;

customer switching;

reduced entry;

market-share changes;

price effects.

The important point is that the simulation can provide economic evidence, not a substitute for the legal test.

11. Case Law 6 — Post Danmark

Post Danmark A/S v Konkurrencerådet

C-209/10, Court of Justice, 27 March 2012

Post Danmark concerned selective pricing and alleged exclusionary effects.

The Court examined whether pricing behaviour by a dominant undertaking could actually restrict competition.

ABM relevance

An ABM can test different pricing scenarios:

Scenario A

Dominant firm maintains normal prices.

Scenario B

Dominant firm offers selective discounts.

Scenario C

Competitors respond with lower prices.

Scenario D

A rival exits.

The simulation can estimate the resulting market trajectory.

However, the legal assessment still requires application of Article 102 and the relevant effects-based analysis.

12. Case Law 7 — Deutsche Telekom

Deutsche Telekom v Commission

C-280/08 P, Court of Justice, 14 October 2010

The case concerned margin squeeze.

A vertically integrated dominant undertaking can potentially abuse its position where the relationship between upstream and downstream prices makes effective competition difficult.

ABM application

An agent-based model could represent:

Upstream price

↓

Downstream competitor's costs

↓

Retail price

↓

Customer switching

↓

Competitor viability

This can help estimate whether a particular pricing structure could make downstream competition unsustainable.

13. Case Law 8 — Google Shopping

Google and Alphabet v Commission

C-48/22 P, Court of Justice, 10 September 2024

The Court upheld the Commission's finding concerning Google's preferential treatment of its own comparison-shopping service.

The judgment is relevant to computational antitrust because digital platforms can use complex algorithms to determine:

rankings;

visibility;

traffic;

access;

competitive exposure.

ABM relevance

An ABM could model:

algorithmic ranking → traffic allocation → merchant participation → rival visibility → market shares.

Again, the model would be an analytical tool rather than the legal infringement itself.

14. What an Antitrust Prediction System Looks Like

A sophisticated system could contain five layers.

Layer 1 — Market Data

Inputs:

prices;

quantities;

market shares;

customer switching;

costs;

entry;

contracts.

Layer 2 — Agent Profiles

Each undertaking receives characteristics:

market power;

costs;

capacity;

strategy;

innovation level.

Layer 3 — Behavioural Rules

Agents decide:

prices;

output;

investment;

advertising;

entry;

acquisition;

switching.

Layer 4 — Interaction Engine

Agents observe:

competitors;

consumers;

market conditions.

Layer 5 — Legal/Economic Output

The system produces:

predicted prices;

market shares;

foreclosure rates;

consumer welfare;

entry probability;

innovation effects.

15. The Biggest Legal Problem: Prediction Is Not Proof

This is the central principle.

Suppose an ABM predicts:

“There is an 80% probability that these firms will coordinate.”

That does not mean:

“The firms have committed an Article 101 infringement.”

Why?

Because legal liability depends on evidence and legal standards, not merely statistical probability.

The model may be affected by:

incorrect assumptions;

incomplete data;

incorrect behavioural rules;

calibration problems;

omitted variables;

coding errors;

unrealistic consumer behaviour.

16. Explainability Problem

A competition authority must be able to explain:

Why did the model produce this result?

Consider an AI system that concludes:

“This merger will cause foreclosure.”

The undertaking may ask:

Which variables produced the result?

Which assumptions were used?

What counterfactual was applied?

How sensitive is the outcome?

What happens if one parameter changes?

A black-box prediction is therefore problematic in high-stakes enforcement.

17. Counterfactual Problem

Antitrust often requires comparing:

Actual world

What happened or is proposed.

versus

Counterfactual world

What would have happened without the alleged conduct.

ABM can construct multiple counterfactuals.

For example:

Merger occurs

versus

Merger does not occur

or:

Algorithmic pricing system

versus

Independent pricing

The quality of the legal-economic conclusion depends heavily on whether the counterfactual is realistic.

The Commission itself recognises the importance of counterfactual analysis in competition proceedings. Its materials on Article 102 investigations explain that market assessment involves identifying the competitive conditions and factors relevant to the alleged conduct. (Competition Policy)

18. Calibration

A model must be calibrated against real-world evidence.

Suppose the real market has:

40% Firm A;

35% Firm B;

25% Firm C.

But the ABM generates:

70% Firm A;

20% Firm B;

10% Firm C.

That suggests the model may not reproduce important market dynamics.

A poorly calibrated model can produce apparently sophisticated but unreliable results.

19. Sensitivity Analysis

A responsible antitrust ABM should ask:

What happens if the assumptions change?

For example:

If the conclusion changes dramatically, the authority should recognise that the prediction is sensitive to assumptions.

20. Algorithmic Collusion

One of the most important applications is algorithmic pricing.

Imagine:

Firm A's algorithm

↔

Firm B's algorithm

Each observes:

competitor price;

historical price;

inventory;

demand.

The algorithms repeatedly optimise profit.

They might independently learn that:

“Aggressive price reductions trigger retaliation.”

Consequently, both maintain high prices.

This produces an important competition-law question:

Can independently designed algorithms produce coordinated outcomes without an express agreement?

ABM is especially useful for studying this question.

But the legal answer depends on the particular conduct and applicable Article 101 principles; an economically coordinated outcome alone does not automatically establish a prohibited agreement.

21. Predictive Merger Analysis

ABM can be used before a merger.

Suppose:

Firm A + Firm B

The model can simulate:

price increases;

output reduction;

entry;

innovation;

customer switching;

product repositioning.

The authority could compare:

Merger scenario

against

No-merger scenario.

This can be particularly useful where firms compete dynamically rather than merely through current prices.

22. Dynamic Competition

Traditional merger analysis may focus on:

Current prices and market shares.

ABM can examine:

What happens over time?

For example:

Year 1: merger

↓

Year 2: competitor loses customers

↓

Year 3: rival reduces R&D

↓

Year 4: innovation slows

↓

Year 5: market becomes concentrated.

This can help investigate theories involving innovation competition.

23. Network Effects

Digital markets can be strongly affected by network effects.

ABM can model:

More users

↓

More data

↓

Better service

↓

More users

This feedback loop may allow one platform to grow rapidly.

A competition authority could simulate whether:

interoperability reduces concentration;

data portability increases switching;

multi-homing prevents lock-in;

entry remains viable.

24. Platform Markets

A platform may have several agent groups:

Consumers

choose platforms.

Advertisers

purchase access.

Publishers

supply content.

Developers

create applications.

Platform

sets rules.

An ABM can simulate the interaction among all of them.

This can be particularly useful for:

self-preferencing;

tying;

interoperability;

access restrictions;

platform fees.

25. Article 102 and Prediction Systems

The Commission's current Article 102 framework confirms that dominance assessment considers market shares together with factors such as entry barriers, buyer power, resources and vertical integration. (Competition Policy)

ABM could therefore assist with:

Dominance analysis

Simulate entry barriers.

Foreclosure

Simulate rival exit.

Pricing

Simulate price effects.

Innovation

Simulate R&D responses.

Vertical integration

Simulate upstream/downstream effects.

But the model does not itself determine whether an undertaking is legally dominant.

26. Evidentiary Challenges

An ABM submitted in antitrust proceedings should ideally disclose:

source data;

model architecture;

assumptions;

equations/rules;

parameter values;

calibration process;

validation;

sensitivity testing;

alternative specifications;

reproducibility procedures.

Without these, the opposing party may have difficulty effectively challenging the evidence.

This matters because Article 102 investigations allow undertakings to receive access to the Commission's investigative file and participate in an oral hearing. (Competition Policy)

27. Procedural Fairness

An AI prediction system creates a new procedural question:

How much of the computational model must be disclosed to the investigated company?

Potentially relevant material could include:

model code;

training data;

datasets;

parameters;

assumptions;

sensitivity results.

But disclosure may conflict with:

trade secrets;

confidential business information;

personal data;

cybersecurity.

A competition authority therefore needs a mechanism that balances transparency with legitimate confidentiality.

28. False Positives

A prediction system might identify a competitive market as dangerous.

Example:

The model predicts 90% coordination.

But actual firms:

compete aggressively;

innovate;

frequently undercut each other.

The model may have assumed overly rational behaviour.

This is a false positive.

29. False Negatives

The opposite can also happen.

The model may predict:

“No substantial foreclosure.”

But the actual platform:

strategically restricts API access;

uses discriminatory algorithms;

imposes exclusionary contracts.

The model may simply have failed to include the relevant mechanism.

Therefore:

Model failure can occur in both directions.

30. Human Oversight

A robust antitrust prediction system should therefore follow:

AI/ABM prediction

↓

Economist review

↓

Legal analysis

↓

Evidence verification

↓

Party submissions

↓

Independent assessment

↓

Final decision

The final legal conclusion should not be generated solely by an automated probability score.

31. Competition Damages

ABM can also assist private litigation.

The European Commission notes that infringements of Articles 101 and 102 can cause higher prices or loss of profits and that EU law provides a framework supporting private damages actions. (Competition Policy)

An ABM could estimate:

overcharge;

lost sales;

lost market share;

reduced output;

consumer harm.

For example:

Observed market

Price = €120

Simulated competitive counterfactual

Price = €100

Potential overcharge estimate:

€20 per unit

But the model must be sufficiently robust to establish the counterfactual and causal relationship.

32. Merger Remedies and ABM

ABM can also test remedies.

Suppose a merger raises concerns.

Possible simulations:

No remedy

Market concentration increases.

Divestiture

New competitor enters.

Access remedy

Rivals obtain necessary data.

Interoperability remedy

Customers can switch.

Behavioural restriction

Self-preferencing prohibited.

The authority can compare predicted outcomes.

33. Important Case-Law Table

34. Direct vs Analogical Cases

It is important to be precise.

More directly relevant to algorithmic/computational conduct

Eturas, C-74/14

because a common computerised booking system was central to the case.

Algorithmically relevant but not ABM cases

T-Mobile Netherlands

Intel

Post Danmark

Deutsche Telekom

Google Shopping

These cases establish competition-law principles that can be modelled using ABM, but the courts did not decide them using modern agent-based antitrust prediction systems.

Therefore, there is currently limited direct judicial case law specifically validating ABM as an antitrust decision-making methodology.

35. Legal Risks of Using ABM

Risk 1 — Black-box decision-making

A party cannot meaningfully challenge an unexplained prediction.

Risk 2 — Model bias

The model designer's assumptions can influence the outcome.

Risk 3 — Data bias

Historical data may reflect previous market distortions.

Risk 4 — Behavioural assumptions

Agents may be unrealistically rational.

Risk 5 — Overconfidence

A probability estimate may be mistaken for legal certainty.

Risk 6 — Counterfactual manipulation

Changing the “but-for” world can dramatically change the result.

Risk 7 — Reproducibility

Different researchers may obtain different results.

Risk 8 — Dynamic instability

Small parameter changes may create radically different market outcomes.

36. Best-Practice Antitrust ABM

A legally defensible system should ideally contain:

1. Transparent assumptions

Every major assumption should be documented.

2. Real-world calibration

Historical market data should be used.

3. Multiple scenarios

Do not rely on a single simulation.

4. Sensitivity testing

Test alternative parameter values.

5. Independent validation

A separate economist should test the model.

6. Legal-human review

Economic predictions must be connected to the actual legal test.

7. Procedural disclosure

The affected undertaking should have an appropriate opportunity to challenge material evidence.

37. Hypothetical Example

Suppose five online retailers use dynamic pricing algorithms.

Each algorithm:

observes competitors;

changes prices every minute;

maximises expected profit.

An authority suspects coordinated pricing.

Traditional investigation

Investigators examine:

communications;

contracts;

algorithm instructions;

pricing records.

ABM investigation

The authority creates five artificial agents.

Each receives the same:

costs;

demand;

price rules;

information.

It runs:

Scenario A: independent algorithms

Scenario B: algorithms observe competitor prices

Scenario C: algorithms receive future price information

Scenario D: one algorithm acts as market leader.

If Scenario B consistently generates higher prices, the model provides economic evidence that information visibility may facilitate coordination.

But the authority must still establish the actual legal theory and evidence concerning the real firms.

38. ABM and the Burden of Proof

A critical principle is:

Prediction is evidence; prediction is not automatically infringement.

An authority should distinguish:

Fact

“The algorithm actually changed prices according to this rule.”

Model assumption

“We assume agents maximise profit.”

Prediction

“Under these assumptions, prices increase by 15%.”

Legal conclusion

“The actual conduct satisfies Article 101/102.”

These are four different propositions.

39. Private Litigation

ABM may become increasingly useful in damages actions.

A claimant might argue:

“Without the alleged cartel, prices would have been lower.”

The defendant might respond:

“The model assumes unrealistic consumer behaviour.”

The court may therefore need competing economic models.

This is consistent with the broader EU framework under which national courts play an important role in applying Articles 101 and 102 and hearing private damages actions. (Competition Policy)

40. Future Applications

Agent-based antitrust systems could potentially be used for:

A. AI pricing

Predict algorithmic coordination.

B. Digital platforms

Model network effects.

C. Merger control

Simulate post-merger market structure.

D. Supply chains

Predict bottleneck exploitation.

E. Advertising

Model auction dynamics.

F. Retail

Simulate dynamic pricing.

G. Energy

Model electricity-market competition.

H. Transport

Model platform competition.

I. Financial markets

Model algorithmic trading interactions.

J. Innovation

Simulate long-term R&D competition.

41. Simple Exam Structure

When answering an examination question, use:

A — Agents

Who are the market participants?

B — Behaviour

How do they make decisions?

C — Competition

How do they interact?

D — Model

What assumptions and data are used?

E — Effects

What happens to prices, output, innovation and entry?

F — Evidence

Can the prediction be independently verified?

G — Law

Does the evidence satisfy Article 101 or 102?

H — Procedural fairness

Can the affected undertaking challenge the model?

42. Conclusion

Agent-Based Modeling for Antitrust Prediction Systems represents a potentially powerful intersection of economics, computational science and competition law.

ABM can model markets as interacting systems rather than treating firms and consumers as static variables. It can therefore be particularly useful for examining:

algorithmic pricing;

coordination;

dynamic competition;

merger effects;

platform markets;

network effects;

foreclosure;

innovation;

damages.

However, the central legal limitation is fundamental:

An ABM can predict an economic outcome, but it cannot by itself establish an infringement of Article 101 or Article 102 TFEU.

The strongest legal foundations come from Eturas for computerised systems and concerted-practice evidence; T-Mobile Netherlands for information exchange; AC-Treuhand for participation in cartel mechanisms; Cartes Bancaires for object/effects analysis; Intel and Post Danmark for effects-based exclusionary analysis; Deutsche Telekom for vertical pricing; and Google Shopping for digital-platform foreclosure and self-preferencing.

EU enforcement procedure also gives undertakings procedural rights, including access to the Commission's investigative file and an oral hearing in Article 102 proceedings. (Competition Policy)

Ultra-basic revision formula

ABM Antitrust System = Agents + Behaviour Rules + Market Data + Interaction + Counterfactual + Simulation + Sensitivity Testing → Economic Prediction → Human/Economic Verification → Legal Assessment under Article 101/102.

Most important distinction:

Simulation ≠ Proof; Prediction ≠ Infringement; Model Output ≠ Legal Judgment.