EU Competition Law, Regulation 1/2003 and ECN+

Substantive law sits in Articles 101 and 102 TFEU. These are Treaty provisions. They were Articles 81 and 82 EC before Lisbon and 85 and 86 EEC originally, so older material uses all three numberings.

Procedural law sits in Regulation (EC) No 1/2003, applicable from 1 May 2004, which replaced Regulation 17/62.

What the 2004 reform, usually called modernisation, actually did: abolished the old notification and clearance system under which firms sent agreements to Brussels and waited; made Article 101(3) directly applicable so firms self-assess; decentralised enforcement to national authorities and national courts; and strengthened the Commission's investigative powers in exchange.

State aid, Articles 107 to 109 TFEU, is a separate branch. Not your area and not ATLANTIS's.

The defendant is an undertaking, not a company. It is a functional concept: any entity engaged in economic activity, whatever its legal form. Under the single economic unit doctrine a parent and its subsidiary can be one undertaking, which is how fines reach 10 percent of group turnover.

Jurisdictional threshold: Articles 101 and 102 apply where conduct may affect trade between Member States. Below that, purely national competition law applies.

Article 101 prohibits agreements, decisions of associations and concerted practices that restrict competition. 101(1) prohibits, 101(2) renders them void, 101(3) exempts where four cumulative conditions are met: efficiency gains, a fair share passed to consumers, indispensability of the restrictions, and no elimination of competition.

Article 102 prohibits abuse of a dominant position. There is no equivalent exemption clause, though objective justification and efficiency defences exist in the case law. Dominance is the ability to behave independently of competitors and customers. Market share around 40 percent becomes indicative; 50 percent and above raises a presumption. Dominant firms carry a special responsibility not to impair competition.

Mergers are separate and ex ante. Notify above turnover thresholds, do not close before clearance — closing early is gun-jumping — then Phase I or an in-depth Phase II. The test is whether the deal would significantly impede effective competition, abbreviated SIEC.

A cartel is coordination between competitors instead of competing. Horizontal means between rivals; vertical means between a supplier and a distributor. Horizontal is the serious kind. The bar for finding coordination is low: a concerted practice requires no formal agreement.

• Price fixing. Agreeing prices, floors, surcharges, discount levels or a shared formula. It need not be the final price. Agreeing not to discount, agreeing on list prices, or agreeing on the timing of increases all count.
• Market sharing. Splitting the market so nobody has to compete, by geography, customer or product line. Includes agreeing not to poach each other's clients.
• Output restriction. Agreeing to limit production or capacity to hold prices up.
• Bid rigging. Collusion in tenders. The next section covers it properly, because it is what most screening tools target.

Three more that matter for a data-driven project:

• Information exchange. Sharing commercially sensitive information — future prices, capacity plans, costs — can infringe Article 101 on its own, with no agreement on price, because it removes the uncertainty that forces firms to compete. This matters because exchange leaves traces in data.
• Hub-and-spoke. Rivals coordinating indirectly through a common supplier, distributor or platform rather than contacting each other. No direct rival-to-rival contact, same effect.
• Algorithmic collusion. Pricing algorithms converging on supra-competitive prices with no human agreement. Genuinely unresolved: the doctrine needs a concurrence of wills, and nobody knows what that means when two reinforcement learners learn to tacitly collude. Schrepel works here, and his podcast has an episode on two language models doing exactly this in a repeated Bertrand duopoly.

Collusion among bidders in a procurement tender. The recognised forms:

• Cover bidding, also called complementary or courtesy bidding. The losers deliberately submit high or defective bids so the designated winner looks competitive.
• Bid suppression. Agreeing to withdraw a bid, or not to bid at all.
• Bid rotation. Taking turns being the winner across a series of tenders.
• Market allocation in tenders. Splitting contracts by region, buyer or contract type.
• Subcontracting as payoff. The winner subcontracts work to the losers as compensation.

Exclusionary abuses foreclose rivals: predatory pricing, exclusive dealing, loyalty rebates conditional on buying most or all requirements, tying and bundling, refusal to supply or to grant access to an essential input, margin squeeze by a vertically integrated firm, self-preferencing on your own platform as in Google Shopping, and interoperability denial.

Exploitative abuses harm trading partners directly: excessive pricing, and unfair trading terms. Meta v Bundeskartellamt was an unfair-terms case.

This is the context almost no other candidate will have, and it reframes computational antitrust from enthusiasm to necessity.

Cartels were historically found through leniency: a participant confesses in exchange for immunity and hands over the evidence. For two decades that was the dominant detection channel in the EU.

Leniency applications have fallen substantially. The standard explanation is private damages exposure: after Directive 2014/104 made follow-on claims easier, confessing to the regulator marks you as a target for claimants. Immunity became more expensive, so fewer firms buy it.

Structural screens look at market characteristics that make collusion likely: few firms, homogeneous products, stable demand, high entry barriers, price transparency, frequent interaction. They tell you where to look, not who did it.

Behavioural screens look for patterns in outcome data that competition should not produce. In bid data:

• Market shares suspiciously stable over time
• Rotation patterns in who wins
• Price variance that is too low — genuine competition is noisy
• Losing bids clustered just above the winning bid
• Winning margins that do not fall as the number of bidders rises, when real competition should push them down
• Identical or near-identical bids, shared typos, sequential document metadata, matching formatting
• Pairs of firms that never bid against each other
• Benford's law tests on bid digits, to detect fabricated numbers
• Network analysis of co-bidding, shared personnel, shared addresses, shared ownership

In price data: parallel movements, simultaneous jumps, suspicious stability, anomalies against cost indices.

From Schrepel's own recruitment post. Know them, and know they are doing four different things.

The classification is the valuable move, because the legal problems differ completely by category:

1. Detection and screening — Spain, Brazil, Chile. These decide who gets investigated. Legal problems: what authorises collection before any investigation exists, selection bias, false positives, and whether the target ever learns a model flagged it.
2. Evidence analysis — Greece. Processing lawfully seized material. Legal problems: rights of defence, access to the file, reliability of the inference, and whether the firm can replicate the analysis.
3. Internal knowledge retrieval — France. Legal problems: retrieval integrity, and whether a wrong retrieval silently contaminates a decision.
4. Market research on third parties — Poland. Closer to consumer protection than to enforcement.

Chapter V of Regulation 1/2003. This is the legal heart of the data problem.

How a case runs, in order: initiation, whether ex officio, on a complaint or from a leniency application; investigation through requests and inspections; Statement of Objections, which is the formal charge and triggers access to the file; written reply and oral hearing; Advisory Committee consultation; decision under Article 7 for infringement and fine, Article 9 for commitments, or rejection; then judicial review.

This is where AI collides with the framework, so know this table better than any other in the chapter.

Standard of proof: the Commission must adduce sufficiently precise and consistent evidence.

Article 23(2): for infringements of Article 101 or 102 committed intentionally or negligently, up to 10 percent of total worldwide turnover in the preceding business year. Note that culpability threshold — it is the identical formula to GDPR Article 83(2).

How the amount is built under the 2006 Guidelines: a basic amount of up to 30 percent of relevant sales multiplied by years of participation; for cartels an additional entry fee of 15 to 25 percent of the value of sales; adjustments for aggravating factors such as recidivism or ringleader role and mitigating ones; a deterrence multiplier; leniency reductions, with full immunity for the first qualifying applicant; a settlement reduction of 10 percent; and an inability-to-pay adjustment in exceptional cases.

• Article 263 TFEU — annulment action before the General Court, with appeal to the Court of Justice on points of law only.
• Article 261 TFEU with Article 31 of Regulation 1/2003 — unlimited jurisdiction over fines. The Court may cancel, reduce or increase them.
• KME Germany and Chalkor, 2011 — the Court must conduct an in-depth review of both law and fact on fines, not defer to the Commission's margin of appreciation.
• Tetra Laval — where the Commission relies on complex economic assessment, its evidence must be convincing.

Directive (EU) 2019/1, adopted 11 December 2018, transposition deadline 4 February 2021.

The problem it solved: Regulation 1/2003 decentralised enforcement to national authorities but did not guarantee they had adequate powers. Some lacked independence, resources, or the ability to impose effective fines, so identical conduct met very different enforcement depending on the Member State.

What it requires: independence from political and commercial influence with security of tenure; adequate human, financial and technical resources; investigative powers mirroring Regulation 1/2003; decision-making powers including interim measures and commitments; effective turnover-based fines; harmonised leniency programmes with a marker system; mutual assistance for notifying and enforcing fines across borders; and harmonised limitation periods.

Substantive cases, lighter touch, for recognition only: Hoffmann-La Roche on dominance and special responsibility, United Brands on dominance, Intel C-413/14 P on rebates and effects analysis, Google Shopping and Google Android on self-preferencing.

This table is the project in one page. If you memorise one artefact from this chapter, memorise this.

ATLANTIS is a five-year project funded by a two million euro ERC Consolidator Grant, running 2026 to 2031, hosted at VU Amsterdam. Schrepel's recruitment post describes three strands: accuracy, fairness, and the institutional arrangements that sustain both. Everything in this chapter maps onto one of them.

Four people on the call. Knowing what each of them publishes tells you what they are listening for.

Schrepel wrote that whether Regulation 1/2003 and the ECN+ Directive need updating, and how, is an open question and, remarkably, an almost unstudied one. That is an unusual admission in a recruitment post, and it tells you the field is open. Six projects below, each with the gap, the method, why it would work, and what it produces.

Every row below is documented on your CV or in a paper, so every row survives a follow-up question. That matters more than breadth.

1. Articles 101 and 102 TFEU prohibit; Regulation 1/2003 enforces; ECN+ extended it to all 27 national authorities.
2. Object versus effect decides how much the agency must prove. Cartel conduct is by object.
3. The four cartel forms are price fixing, market sharing, output restriction and bid rigging, plus information exchange, hub-and-spoke and algorithmic collusion.
4. Bid rigging is the screening target because procurement data is structured and often public.
5. Leniency has declined, which is why agencies now screen proactively.
6. Screens produce flags, not evidence, and that gap is where law meets machine learning.
7. The five agency tools do four different jobs with four different legal problems.
8. Article 18 necessity plus HeidelbergCement is the data problem.
9. Article 296 reasons plus Article 27 access to the file is the fairness problem.
10. Article 23(5) says not criminal; Menarini says criminal. Use that gap.