Adversarial Robustness of Economic Models Used in Enforcement

1. Meaning

Adversarial robustness of economic models used in enforcement means the ability of an economic model, econometric analysis, algorithm, or quantitative test used by a regulator, competition authority, financial authority, or court to produce reliable conclusions even when the parties being investigated deliberately exploit weaknesses in the model.

In enforcement proceedings, economic models may be used for:

market definition;

market-power assessment;

price-cost tests;

predatory pricing;

margin squeeze;

cartel overcharge estimation;

damages calculation;

merger simulation;

counterfactual analysis;

effects of exclusionary conduct;

pass-on analysis;

algorithmic pricing investigations;

assessment of barriers to entry;

calculation of foreclosure or market shares.

An adversarial environment is different from ordinary statistical modelling. A regulated undertaking knows that the authority is testing a particular economic proposition and may therefore:

select data strategically;

change prices or output temporarily;

alter algorithms;

exploit model assumptions;

provide incomplete information;

target variables that the model does not observe;

create behaviour that appears compliant while preserving the underlying exclusionary effect.

Therefore, an enforcement model should not merely fit historical data. It should also be robust to strategic behaviour and alternative plausible assumptions.

2. Why Economic Models Matter in Enforcement

Modern enforcement increasingly relies on quantitative evidence.

For example, suppose a competition authority investigates whether a dominant platform excluded competitors.

The authority might construct:

Observed outcome = competitive counterfactual + effect of allegedly unlawful conduct + other factors

The central difficulty is that the counterfactual is unobservable.

The authority must estimate what would have happened if the conduct had not occurred.

This creates several possible vulnerabilities:

Adversarial robustness therefore becomes an important part of procedural fairness and evidentiary reliability.

3. European Legal Framework

There is no single EU statute called an "Adversarial Economic Model Robustness Act."

Instead, the concept emerges from several principles:

A. Article 102 TFEU

For abuse of dominance, authorities may need to establish:

dominance;

conduct;

exclusionary or exploitative effect;

causal connection;

competitive harm;

sometimes consumer harm.

Economic evidence can be central to these questions.

B. Article 101 TFEU

Economic analysis may be used to determine:

restrictive effects;

market foreclosure;

cartel effects;

efficiencies;

counterfactual competitive conditions.

C. EU merger control

Economic modelling can be used in:

unilateral-effects analysis;

coordinated-effects analysis;

efficiencies;

diversion ratios;

upward pricing pressure;

entry analysis.

D. Digital Markets Act

The DMA creates additional quantitative and data-intensive enforcement problems involving:

gatekeepers;

interoperability;

self-preferencing;

data access;

ranking;

switching;

advertising transparency.

E. EU procedural principles

Economic models used against undertakings must operate within broader requirements of:

rights of defence;

statement of objections;

access to evidence;

sufficient reasoning;

judicial review;

proportionality;

reliable evidence.

Thus, a technically sophisticated model does not automatically produce a legally sufficient enforcement decision.

4. What Makes an Economic Model "Adversarially Robust"?

A robust enforcement model should survive several forms of challenge.

4.1 Specification robustness

The authority should test whether the conclusion survives reasonable alternative specifications.

For example:

Model A

Demand = price + income + advertising

should be compared with plausible alternatives involving:

price + income + advertising + seasonality + competitor entry.

If the alleged anticompetitive effect disappears merely because one reasonable variable is added, the original conclusion requires careful qualification.

5. Data Robustness

Economic models are only as reliable as their underlying data.

Important questions include:

Who supplied the data?

Was it complete?

Were missing observations excluded?

Were outliers removed?

Were internal transactions included?

Was data generated after the investigation began?

Are competitor data available?

Are algorithmically generated observations independent?

Are prices net or gross of discounts?

An authority should ideally perform:

Data validation → cleaning → independent verification → sensitivity analysis → model estimation.

6. Counterfactual Robustness

This is one of the most important issues.

Suppose a regulator claims:

Company X's conduct caused prices to increase by 15%.

The critical question is:

15% compared with what?

Possible counterfactuals include:

prices before the conduct;

prices in another geographical market;

prices charged by competitors;

prices predicted by a structural demand model;

prices under a simulated competitive equilibrium.

Each may produce a different result.

Therefore, enforcement should test several plausible counterfactuals rather than treating one model as unquestionably correct.

7. Sensitivity Analysis

A model should be tested against reasonable changes in assumptions.

For example:

If the legal conclusion remains substantially unchanged, confidence in the model increases.

If the result changes radically, the authority should disclose that uncertainty.

8. Adversarial Manipulation of Models

An undertaking may potentially respond strategically once it understands the enforcement methodology.

Examples:

8.1 Gaming a price test

If an authority examines whether:

Price < Average Avoidable Cost

a firm might temporarily increase prices during the period examined.

8.2 Gaming market-share calculations

A firm might restructure transactions among affiliated entities.

8.3 Gaming ranking algorithms

A platform might modify ranking signals so that prohibited self-preferencing becomes difficult to detect.

8.4 Gaming an econometric benchmark

A firm might deliberately introduce temporary volatility, making a clean counterfactual harder to estimate.

This means enforcement methodology should not be entirely predictable where predictability enables circumvention.

However, secrecy cannot eliminate the undertaking's rights of defence. The authority must still provide sufficient information for meaningful challenge.

9. Case Law

Case 1 — Airtours v Commission

Airtours plc v Commission, T-342/99, General Court, 6 June 2002

Facts

The Commission prohibited the proposed acquisition involving Airtours on the basis that the merger would create conditions conducive to coordinated effects.

The Commission relied substantially on an economic theory concerning tacit coordination.

Importance

The General Court subjected the Commission's economic reasoning to detailed judicial scrutiny.

The Court concluded that the Commission had not established the necessary conditions with sufficient evidence.

Relevance to adversarial robustness

This case demonstrates that:

An economic theory is not enough; its underlying assumptions must be supported by evidence.

An enforcement model must demonstrate that the relevant economic conditions actually exist.

10. Case 2 — Tetra Laval v Commission

Commission v Tetra Laval, C-12/03 P, CJEU, 15 February 2005

Facts

The Commission prohibited Tetra Laval's acquisition of Sidel, relying partly on a theory of future leveraging and conglomerate effects.

The European courts examined whether the Commission had adequately established the predicted economic effects.

Principle

The Court emphasized the need for sufficiently convincing evidence when enforcement depends upon predictions concerning future economic behaviour.

Relevance

This is particularly important for adversarial models because predictions may be sensitive to assumptions about:

future entry;

customer behaviour;

competitor response;

incentives;

technological change.

A model predicting future harm should therefore be stress-tested against alternative behavioural assumptions.

11. Case 3 — Microsoft v Commission

Microsoft Corp. v Commission, T-201/04, General Court, 17 September 2007

Facts

The Commission found abuses involving interoperability information and tying of Windows Media Player.

The case involved complex technological and economic questions concerning interoperability and competition.

Principle

The Court carefully examined:

market effects;

competitive foreclosure;

technological circumstances;

Microsoft's justifications;

evidence concerning competitors.

Relevance

The case illustrates the difficulty of evaluating conduct in technologically dynamic markets.

An enforcement model that assumes competitors behave in a static manner may be unreliable where:

technology changes rapidly;

products evolve;

network effects exist;

interoperability changes competitive conditions.

12. Case 4 — Intel

Intel Corp. v Commission, C-413/14 P, CJEU, 6 September 2017

Facts

Intel was fined for practices involving rebates granted to major computer manufacturers and a retailer.

The Commission treated the rebates as abusive because of their capability to foreclose competitors.

Important development

The CJEU held that where the dominant undertaking submits evidence that its conduct was not capable of restricting competition, the Commission must examine all relevant circumstances, including the as-efficient-competitor (AEC) test, where appropriate.

Relevance

This is highly important for model robustness.

An enforcement authority cannot necessarily rely on a simplified theoretical assumption if quantitative evidence concerning actual foreclosure capability is relevant.

The economic analysis should consider:

dominant firm's market position;

market coverage;

duration;

rebate conditions;

foreclosure share;

AEC analysis;

strategy and commercial context.

13. Case 5 — Intel Reconsideration

Intel Corp. v Commission, C-413/14 P, followed by Commission reassessment and General Court proceedings

The Intel litigation demonstrates another important aspect of adversarial robustness:

Economic evidence may need to be reassessed when the legal test requires closer examination of actual or potential exclusionary effects.

This makes the model itself part of the evidentiary process rather than an unquestionable mathematical conclusion.

14. Case 6 — Qualcomm

Qualcomm Inc. v Commission, T-235/18, General Court, 15 June 2022

Facts

The Commission imposed a fine on Qualcomm concerning payments to Apple.

The Commission used an economic analysis including an as-efficient-competitor type price-cost assessment to examine foreclosure.

General Court's approach

The Court identified serious problems concerning the Commission's analysis, including:

insufficient examination of certain relevant circumstances;

errors concerning the economic assessment;

procedural problems affecting the Commission's analysis.

The decision was annulled.

Relevance

This is one of the clearest modern European examples showing why an economic enforcement model must be internally consistent, complete and capable of adversarial testing.

A sophisticated quantitative test can still fail if:

important evidence is omitted;

assumptions are unsupported;

the test does not correspond properly to the legal theory of harm.

15. Case 7 — Google Shopping

Google and Alphabet v Commission, C-48/22 P, CJEU, 10 September 2024

Facts

The case concerned Google's treatment of its own comparison-shopping service in general search results.

The Commission found that Google had positioned and displayed its own comparison-shopping service more favourably than competing services.

Importance

The CJEU upheld the essential finding of abuse.

The case is particularly significant for digital enforcement because it concerns:

algorithmic ranking;

self-preferencing;

platform dominance;

traffic diversion;

actual and potential competitive effects.

Relevance to adversarial robustness

Digital enforcement models must account for:

dynamic algorithms;

changing ranking systems;

user responses;

competitor adaptation;

traffic diversion;

network effects.

A static model may miss these interactions.

16. Case 8 — Servizio Elettrico Nazionale

Servizio Elettrico Nazionale SpA and Others, C-377/20, CJEU, 12 May 2022

Principle

The CJEU examined exclusionary conduct under Article 102 TFEU and clarified that conduct may be abusive where it is capable of restricting competition, rather than requiring proof that competitors were completely eliminated.

Relevance

This is important for economic modelling because the model should distinguish between:

actual observed harm

and

capacity of conduct to restrict competition.

An enforcement model therefore must be designed around the correct legal theory of harm rather than simply searching for a large observed price effect.

17. Case 9 — Deutsche Telekom

Deutsche Telekom AG v Commission, C-152/19 P, CJEU, 25 March 2021

Importance

The CJEU considered exclusionary conduct involving access conditions and margin squeeze.

The case clarified the relationship between different theories of exclusion and the conditions under which the Bronner test applies.

Relevance

The lesson for quantitative enforcement is:

The economic test must correspond to the legal characterization of the conduct.

For example, an authority should not automatically apply the economic methodology appropriate to a refusal-to-deal case to a completely different exclusionary practice.

18. Case 10 — Slovak Telekom

Slovak Telekom a.s. v Commission, C-165/19 P, CJEU, 25 March 2021

The case concerned access conditions and exclusionary behaviour in telecommunications.

The CJEU distinguished the legal treatment of different forms of exclusionary conduct.

Economic relevance

Telecommunications markets demonstrate why models must consider:

network economics;

access costs;

downstream competition;

regulated inputs;

wholesale and retail prices;

economies of scale.

An apparently simple price-cost model may therefore fail if it ignores the structure of the regulated network.

19. Case 11 — Google Android

Google and Alphabet v Commission, C-738/22 P, CJEU, 2 July 2026

The case concerns Google's Android ecosystem and restrictions involving:

search;

mobile operating systems;

app distribution;

contractual arrangements;

competitive foreclosure.

It demonstrates the increasing importance of ecosystem-level economic analysis.

A model examining one contractual restriction in isolation may fail to capture:

operating-system effects + app distribution + search defaults + network effects + user switching.

Thus, enforcement models increasingly need to account for interactions among several interconnected markets.

20. Case 12 — Google AdSense for Search

Google and Alphabet v Commission, T-334/19, General Court, 18 September 2024

This is particularly relevant to advertising economics.

The case concerned Google's conduct in online search advertising intermediation.

The General Court annulled the Commission decision because of errors in the assessment of the duration and coverage of the relevant contractual clauses and their competitive effects.

Importance

It demonstrates a fundamental principle:

Economic evidence must correspond precisely to the conduct, period and market actually established in the case.

An enforcement model may be statistically sophisticated but legally inadequate if its observations do not match the precise infringement period or contractual mechanism.

21. Lessons From the Cases

22. Adversarial Testing Framework

A European enforcement authority can conceptually subject an economic model to five levels of testing.

Level 1 — Data test

Ask:

Is the underlying dataset complete, accurate and independently verifiable?

Level 2 — Specification test

Ask:

Does the conclusion survive reasonable alternative models?

Level 3 — Counterfactual test

Ask:

Does the result survive reasonable alternative counterfactuals?

Level 4 — Strategic-response test

Ask:

Could the investigated undertaking or its competitors strategically alter behaviour to defeat the model?

Level 5 — Legal-fit test

Ask:

Does the economic result actually establish the elements of the relevant legal infringement?

The fifth level is crucial.

A statistically significant coefficient does not automatically equal an Article 101 or Article 102 infringement.

23. Machine-Learning Models and Enforcement

AI-based enforcement creates additional robustness problems.

Suppose an authority uses a machine-learning model:

Probability of cartel behaviour = 0.87

That number alone does not establish liability.

Potential problems include:

training-data bias;

adversarial examples;

distribution shift;

hidden variables;

data leakage;

overfitting;

false positives;

false negatives;

correlation without causation;

opaque feature engineering.

A company might also deliberately alter behaviour to move its conduct outside the model's learned distribution.

Therefore:

Prediction ≠ proof of infringement.

The model may be an investigative tool, but the final legal conclusion requires appropriate evidence under the applicable legal standard.

24. Economic Models and Algorithmic Pricing

This issue becomes particularly important with algorithmic pricing.

Imagine four competitors use pricing algorithms.

The algorithms independently learn:

"Prices of competitors are high → increase own price."

No human employee explicitly agrees to fix prices.

An enforcement authority may use:

price correlation;

variance analysis;

structural-break tests;

event studies;

algorithm logs;

communication records;

simulations.

But an adversarially robust analysis must distinguish:

parallel pricing

from

coordination

and from

unilateral algorithmic adaptation.

The economic model must therefore be integrated with documentary, technical and behavioural evidence.

25. Model Transparency and Rights of Defence

A major legal issue is the balance between:

Regulatory confidentiality

Authorities may need to protect:

detection techniques;

confidential algorithms;

investigative methods.

versus

Defendant's rights

The undertaking must have a meaningful opportunity to challenge:

data;

assumptions;

methodology;

statistical significance;

counterfactual;

causal inference.

This produces an important principle:

A model should not become a black box that the defendant cannot meaningfully contest.

26. Confidential Algorithms and Economic Evidence

Suppose a competition authority uses an algorithm that calculates:

foreclosure probability = 74%.

The undertaking should be able, subject to legitimate confidentiality protections, to understand sufficiently:

what variables were used;

what period was examined;

what assumptions were made;

how missing data were handled;

how the model was validated;

what alternative specifications were tested;

how uncertainty was calculated.

Otherwise, judicial review becomes difficult.

27. Robustness and Standard of Proof

Economic evidence does not operate in isolation.

The authority generally needs to establish the legally relevant facts according to the applicable evidentiary standard.

Therefore:

Strong model + weak underlying evidence

= potentially insufficient.

Weak model + strong documentary evidence

= model may be unnecessary or merely supplementary.

Strong model + strong documentary evidence

= quantitatively and legally more persuasive evidence.

The model is therefore part of an evidence ecosystem.

28. Civil Damages Claims

Adversarial robustness is equally important in private damages litigation.

Suppose a cartel victim claims:

"I paid €100 million more because of the cartel."

The court may need to estimate:

Actual price − hypothetical competitive price = overcharge

But the hypothetical competitive price is unobservable.

Experts may use:

before-and-after analysis;

difference-in-differences;

comparator markets;

regression analysis;

synthetic controls;

structural models.

The defendant can challenge:

comparator selection;

time period;

demand shocks;

cost changes;

inflation;

market structure;

pass-on assumptions.

Thus, robustness is central to accurate damages.

29. Practical Checklist for Courts and Regulators

Before relying heavily on an economic model, ask:

Data

Is the dataset complete?

Are there missing observations?

Are there selection problems?

Are the variables independently verifiable?

Methodology

Why was this model selected?

Were alternative models tested?

Are the assumptions economically justified?

Is there overfitting?

Causation

Does correlation establish causation?

What is the counterfactual?

Were alternative explanations examined?

Adversarial robustness

Could the undertaking manipulate the relevant variables?

Could competitors strategically respond?

Could the algorithm adapt?

Does the model remain valid outside its training period?

Legal relevance

Does the model address the actual legal test?

Does it establish harm, capability of harm, or merely correlation?

Is the relevant period correct?

Is the relevant market correctly identified?

Procedural fairness

Can the undertaking meaningfully challenge the model?

30. Key Distinction: Robustness vs Accuracy

These concepts should not be confused.

Accuracy asks:

"Does the model correctly predict the underlying phenomenon?"

Robustness asks:

"Does the conclusion remain reliable when assumptions, data, behaviour or circumstances change?"

A model may be highly accurate in ordinary circumstances but poorly robust against strategic manipulation.

For enforcement, both qualities matter.

31. Remedies Where Models Are Unreliable

If a quantitative model proves unreliable, an authority or court may need to:

obtain additional data;

revise the model;

perform sensitivity analysis;

use alternative methodologies;

narrow the infringement theory;

rely on documentary evidence;

commission an independent expert;

conduct further investigation;

reconsider the counterfactual.

A model should not be treated as a substitute for legal reasoning.

32. Exam-Ready Conclusion

Adversarial robustness of economic models used in enforcement concerns whether quantitative and econometric tools remain reliable when their assumptions, data, counterfactuals and strategic vulnerabilities are challenged by the investigated undertaking.

European case law, particularly Airtours, Tetra Laval, Intel, Qualcomm, Google Shopping, Servizio Elettrico Nazionale, Deutsche Telekom, Slovak Telekom and Google AdSense, demonstrates that economic analysis must be connected to the precise legal theory of harm and supported by sufficiently reliable evidence.

The central principles are:

A model is evidence, not the legal conclusion itself.

Alternative reasonable specifications should be tested.

Counterfactual assumptions must be transparent and defensible.

Strategic manipulation and algorithmic adaptation must be considered.

Statistical correlation should not automatically be treated as causation.

The model must correspond to the precise conduct, market and period under investigation.

Rights of defence require meaningful ability to challenge important economic evidence.

Judicial review can scrutinize whether economic assumptions and evidence adequately support the enforcement conclusion.

Ultra-short rule

In European enforcement, an economic model is sufficiently robust only when its data, assumptions, methodology, counterfactual and causal conclusions withstand reasonable alternative explanations and adversarial scrutiny, while remaining properly connected to the applicable legal test.