earino/ec-merger-phase2-opening

Dataset Factory · v2026.10

Dataset website · GitHub release · Hugging Face · Catalogue

Artifact: e96929efbbc4fcb84d2986a7e32f83f9c5fec5395124439fb3404faff2e9e662

Suitability and evaluation requirements

Retained for retrospective model-development research. Every shipped feature is a same-row case attribute fixed at notification (an administrative record), and eligibility reads only notification-time fields, but the register publishes a single current snapshot with no historical vintages, so training-label timing and full-row historical availability are unverified and no operational or point-in-time forecasting claim is supported.

Supported uses: Retrospective train/validation model-development experiments under the documented timing, vintage and selection limitations; Final evaluation only under a protocol that isolates the labeled holdout from agent development and prevents any later row influencing an earlier prediction

Not supported: Claims of verified real-time, operational or point-in-time forecasting validity; Treating development/validation (eval-split) scores as untouched final-test results; Presenting current-vintage register values (notably the NACE Rev 2 sector coding) as the values an analyst at notification time would have seen

Evaluation access: isolated rows. The labeled eval split is development/validation data. Keep the final holdout outside agent development; prevent later rows from influencing earlier predictions.

Historical availability: partial. Training-label timing: unverified. Vintage: current snapshot. Selection: prediction time.

Cross-row audit: bounded no recovery. Gate 1.5.0 executed its two declared probes on the full artifact on the worker across all three splits (unique case_number donor; repeat case_title with a 365-day offset): 0 rows covered, 0 mismatches, 0 ambiguous on both probes and all splits. Rows are unique notified cases and every feature is a same-row attribute, so no outcome-bearing cross-row exposure is constructed; the probes are a bounded mechanism, not a proof, and no broader attack search was run.

Structural qualification and baseline scores do not certify operational forecasting validity.

• The register retains no historical vintages (bulk feed since 2025-03-18; S3 objects expire after 30 days), so a reverse-history probe is not performable: this is retrospective research over the pinned 2026-10-27 snapshot, never a certified historical forecasting setup.
• Feature availability is fixture-verified at three cross-era points (M.877 1997, M.4731 2007 with its OJ prior notice, M.10188 2021) but not verified row-by-row against notification-era vintages; the sector taxonomy labels are current-vintage NACE Rev 2 applied across all eras.
• Training-label timing is unverified: labels are the register's current values, observed as of the fetch snapshot, with no availability ledger.
• Positives cluster by enforcement era and sector acquisition waves beyond the day-level event key; per-notification-year positive counts (1-22, median about 8) are reported beside every score.
• Prevalence falls across splits (train 4.01%, eval 2.60%, holdout 1.58%): the eval and holdout windows are prevalence-shifted relative to training, an out-of-distribution property reported, not hidden.
• Pending cases with neither an opener nor a terminal decision at fetch (33, overwhelmingly live 2020s filings) are censored and counted per split and year; their labels are unobserved, not negative, and censoring never decides another row's membership.
• The ever-adopted label keeps M.10188 positive despite its 2024-09-06 post-ECJ withdrawal of the opener; M.3333 stays positive on its 2007-03-01 opener with the 2004 pre-notification opener excluded by rule.
• The stock runner baseline trails the required sector-division comparator on the identical eval rows (0.6973 vs 0.7679; paired difference -0.0706, 95% CI [-0.1558, +0.0168], inconclusive with the observed gap negative). Headroom above that prior is NOT demonstrated by the packaged evidence; both numbers must be quoted together.

Corrections to the historical documentation:

• 2026-10-27/28 descriptor correction per the independent pre-flip audit (PUBLISH WITH CONDITIONS): meta.json and quality.json regenerated (retrospective-use honesty, metadata contradiction fixes, recipes shipped, repeated-transaction dependence disclosed); superseded artifact e5a98dfc9c05d0b0..., CSVs byte-identical; re-qualified as e96929ef by worker job ecm-requal-002 before any public visibility

Review: 2026-10-28; applies to v2026.10 / e96929efbbc4fcb84d2986a7e32f83f9c5fec5395124439fb3404faff2e9e662. Frozen data and tags are unchanged.

EU Merger Phase II Opening (v2026.10)

Status: public. Released after explicit operator authorisation; publication changed no data file, tag or checksum.

A concentration is notified to the European Commission under the EU Merger Regulation. The Commission's case file is registered that day with a fixed set of attributes: the merging parties (the case title), the NACE sectors of their activities, the regulation in force, and the notification date itself. At the notification date, from those registration-day attributes only, will the Commission open an in-depth Phase II investigation (adopt an Art. 6(1)(c) decision) in this case? One row per notified merger case; the label is the procedure outcome, which unfolds over the following weeks (302 of 309 openers arrive within two months of notification).

Everything the register records during or after review is excluded by construction and never ships: the decisions block (the label source, including all Art. 6(1)(a)/(b) and Art. 8 outcomes and their adoption dates), caseInvestigationPhase, deadlines, last-decision dates, court cases, renotification and initiation dates, attachments, and caseSimplified (a simplified filing upgraded to Normal during review would make its current value post-hoc). The shipped features are the registration-day facts: case_title, sector_divisions (the case's NACE divisions), sector_count, has_sectors, party_count, title_has_jv, regulation, and calendar decompositions of the notification date. The qualification gate verifies over the shipped columns that no column determines the label and that the label source is absent from every file.

• Artifact version: e96929efbbc4fcb84d2986a7e32f83f9c5fec5395124439fb3404faff2e9e662
• Version tag: v2026.10 (use this revision for reproducibility)
• Rows: 10,091 notified merger cases across three disjoint temporal splits
• GitHub counterpart: https://github.com/earino/ec-merger-phase2-opening

Splits

split	rows	notification years	positive rate
train	5,406	1990-2013	4.01%
eval	1,770	2014-2018	2.60%
holdout	2,915	2019-2026 (snapshot)	1.58%

The splits are temporal on the notification year and disjoint, with both classes in every window and every year (309 positives overall: 217 / 46 / 46). Positives cluster by enforcement era and sector acquisition waves, so the gate's clustered unit is the notification day (45 distinct positive dates in eval, 44 in holdout) and per-notification-year counts (1-22, median about 8) are reported beside every score. The positive rate falls 4.01% -> 2.60% -> 1.58% across the splits - a prevalence shift recorded as a property of the enforcement regime, so the eval and holdout windows are partly out-of-distribution tests. 33 pending cases (no opener and no terminal decision at the snapshot, overwhelmingly live 2020s filings) are censored and counted per split, never labelled: their outcomes are unobserved, not negative.

Reference points

Measured on the identical 1,770 eval rows, after gate acceptance, through the benchmark's unmodified runner (train.py + validate.sh, digests verbatim):

Reference / model	Eval AUC	Brier
Full model (stock XGBoost baseline)	0.6973	0.025343
Sector-division prior (train rate per NACE division) - the required comparator	0.7679	0.025008
Sector-combination prior (train rate per exact division set)	0.7051	0.026192
Title-token prior (train rate per party-name token)	0.6644	0.026111
Calendar/identity prior (train rate per regulation x month)	0.5606	0.025295
Recent-era constant (last three train years)	0.5000	0.025314

Paired row-bootstrap differences (1000 resamples): full model - sector prior -0.0706 [-0.1558, +0.0168] - the interval contains zero and the observed gap is negative: the stock baseline trails the sector prior, recorded as a finding rather than repaired. The model beats the calendar/identity prior (+0.1367 [0.0419, 0.2321]) and the recent-era constant (+0.1973 [0.1068, 0.2745]), so the task is not solved by calendar and identity alone; against title tokens the difference is inconclusive (+0.0329 [-0.0770, +0.1367]). Majority-class accuracy on this split is 0.9740 at a 2.60% base rate - uninformative; AUC is the metric. Persistence is inapplicable: each row is a newly notified case with no prior same-entity observation at the instant. Per-year eval AUC (7-12 positives per year, noisy) is reported in RELEASE_NOTES.md.

Headroom above the sector prior is not demonstrated by this package; equally, one stock model trailing it does not establish that none can beat it. The holdout was never scored by anything and stays sealed; the benchmark's own score is holdout AUC.

Timing, honestly stated

The prediction instant is the notification event itself (an administrative record: the file is registered with these attributes on that day, date-only, Brussels local), so no publication-lag observations apply to the features. The label event - an Art. 6(1)(c) opener - is adopted after notification by construction, and the one pre-notification opener in the register (M.3333, a re-notification carrying its annulled procedure's 2004 opener) is excluded from the label by rule. The register publishes a single current snapshot (bulk feed since 2025-03-18; no historical vintages are retained), so the evaluation contract is retrospective research over the pinned vintage: feature values are stated as current-vintage register values, with notification-era fidelity fixture-verified at three cross-era cases (M.877 1997 against the Commission's own decision PDF, M.4731 2007 against its OJ prior notice, M.10188 2021 against press releases) - not a verified point-in-time reconstruction, and the sector taxonomy is the register's current-vintage NACE Rev 2 coding across all eras.

Source and provenance

• European Commission, Directorate-General for Competition - "EU Competition: Merger case publications" open data (machine-readable case register, coverage from November 1990, updated daily), fetched as one anonymous bulk GET (case-data-M.json, 42,034,960 B, pinned 2026-10-27 vintage; sha256 in MANIFEST.json).
• Construction: code/build.py, executed by worker job ec-merger-003; qualification gate 1.5.0, 46/46 checks on the full artifact, exact-byte acceptance before any figure was computed; baseline job ec-merger-baseline-001 through the unmodified runner contract. Two earlier attempts (ec-merger-001, ec-merger-002) are preserved as failure records.
• Known-event fixtures verified in-file: M.877 BOEING/McDONNELL DOUGLAS, M.4731 GOOGLE/DOUBLECLICK and M.10188 ILLUMINA/GRAIL all reproduce external notification-era sources exactly.

Licence

The source is Commission-owned content under CC BY 4.0 (European Commission reuse notice; DG COMP released the machine-readable case data under it for any processing with acknowledgement). Attribution carries through to every copy of the derived data. Our contribution to the derived compilation is dedicated CC0-1.0; our code is MIT. This is a derived research dataset, not the Commission's official product, and the Commission has not reviewed or endorsed it. See LICENSE.md for the full terms, attribution and citation.

Use with the runner

python3 get_dataset.py --dest ./task
python3 code/qualify_dataset.py ./task
sh baseline/reproduce_baseline.sh ./task