CORDIS — EU research funding (Horizon 2020 & Horizon Europe)

Every research project the European Union has funded since 2014: grant amount, participants, countries, coordinator, dates, topic classification, funding scheme.

• Source: https://cordis.europa.eu/about/archives
• Bulk downloads (no auth, no key):
  • H2020 2014–2020: https://cordis.europa.eu/data/cordis-h2020projects-csv.zip — 55 MB
  • Horizon Europe 2021–2027: https://cordis.europa.eu/data/cordis-HORIZONprojects-csv.zip — 37 MB
• Licence: Creative Commons Attribution 4.0 International for the CORDIS datasets, as recorded on data.europa.eu. Some sibling CORDIS distributions carry the European Commission reuse notice instead. Verified 2026-08-07.
• Attribution: Source: CORDIS, European Commission — cordis.europa.eu (CC BY 4.0)
• Publisher: Publications Office of the European Union

Check the licence field on the specific distribution you download. The catalogue lists CC BY 4.0 and the EC reuse notice against different files in the same dataset family. Both permit reuse with attribution; only one is a standard licence.

Shape

Each zip unpacks to several CSVs, semicolon-delimited, UTF-8:

project.csv columns: id, acronym, status, title, startDate, endDate, totalCost, ecMaxContribution, topics, ecSignatureDate, frameworkProgramme, masterCall, subCall, fundingScheme, nature, objective, rcn, grantDoi, keywords, legalBasis.

organization.csv gives one row per participation: projectID, name, country, nutsCode, activityType, SME, role (coordinator / participant / thirdParty / partner), ecContribution, netEcContribution, totalCost.

Headline figures, computed 2026-08-07:

Why it suits the jam

Money is the most persuasive quantity there is. A number in euros with nine digits reads as fact. Nobody in the room will re-derive it, and nobody will ask which of the four plausible definitions of “the amount” was used — but there really are four, and they disagree by billions.

Every trap here is a definitional one. No sentinel values, no missing-data collapse, no instrument drift. The deceptions are all of the form “which column counts as the money, and which year does it belong to”, which makes it the best complement in the folder to the sensor datasets. It is also the closest to how funding arguments actually go wrong in public.

It is a research-funding dataset at an event about how research misleads, which is a joke the room will enjoy.

The traps

1. totalCost versus ecMaxContribution — a 22% choice

ecMaxContribution is what the EU pays. totalCost is the whole project budget including co-funding from industry and national sources. For H2020:

Report totalCost and “EU research funding” grows by 21.8%, or about €15 bn, without a single dishonest keystroke. Both columns are correctly labelled in a documented schema. Neither is a mistake. The word “funding” in your sentence is doing all the work.

It gets better, because the ratio is not constant across funding schemes:

ERC grants are funded at 100%, so the two columns are identical. Innovation actions require industrial co-funding, so they inflate by 44%. Which measure you pick therefore silently reweights the comparison between fields, favouring applied and industry-linked areas over basic research. A claim of the form “applied research receives more EU money than fundamental research” can be made true or false by choosing a column, and the choice is one word in a groupby.

2. The missing-totalCost artefact — a fake collapse in co-funding

This is the strongest trap in the dataset, and it is not documented anywhere.

totalCost is present on 98.5% of H2020 projects but on only 46.6% of Horizon Europe projects (10 926 of 23 451).

So:

Take the naive route — sum both columns, divide — and Horizon Europe appears to have almost no co-funding at all. The headline writes itself:

“Industrial co-investment in EU research has collapsed: under Horizon 2020 every euro of EU money attracted 22 cents of co-funding; under Horizon Europe it attracts one cent.”

That is fully reproducible, sourced to the Commission’s own open data, and completely false. On the projects where the figure was actually recorded, co-funding is higher than under H2020 (1.272 versus 1.218). The entire “collapse” is a numerator summed over a subset and a denominator summed over everything.

This is the coverage-collapse structure from the fog data (hypotheses/fog-is-disappearing.md) transplanted into finance, and it is nastier here because the missing values are silently absent rather than encoded as a sentinel — there is no -1 to notice.

3. Which year does a grant belong to? — same money, two stories

Every project has a startDate, an endDate and an ecSignatureDate. Attributing the same €130 bn by start year versus signature year:

By start date, 2021→2022 is flat: 10.46 → 10.75, up 3%. By signature date, 2021→2022 is 6.90 → 16.69, up 142%.

Same grants. Same euros. Same file. One column swap, and “EU research funding more than doubled” and “EU research funding was flat” are both true, both reproducible, and both defensible — signature date is when the Commission committed, start date is when work began, and either is a legitimate basis for an annual series.

There is no correct answer to point at. That makes it a better claim to defend than a trap with a right answer, because the Offensio side cannot win by correcting you. It has to argue about meaning, in public, which is exactly the argument the jam wants to stage.

4. Endpoint truncation at both ends

By start date, 2014 shows €1.60 bn and 2027 shows €1.00 bn. Both are artefacts: H2020 only began disbursing in late 2014, and Horizon Europe projects starting in 2027 have mostly not been signed yet. Include either endpoint in a trend line and you can manufacture a boom or a bust. Both endpoints are in the data, so including them requires no filter step at all — the deception is in the absence of a filter, which is much harder to point at than its presence.

5. Coordinator versus participant — who “got” the money

organization.csv records ecContribution per participation. Two defensible ways to say what a country received:

• participant attribution: sum each organisation’s own ecContribution
• coordinator attribution: credit the whole project to the coordinator’s country

Spain gains 27% and Greece 29% by being credited as coordinators; the UK loses 11%. National league tables of research success are published on exactly this choice. Selective agreement — the normaliser deciding what “a country’s funding” means — with a national newspaper story attached.

Austria: €1.96 bn or €2.11 bn, 2.88% or 3.09% of the EU total. Pick the one that suits the argument.

6. Further handles

• euroSciVoc.csv is multi-label. Projects carry several topic classifications, so summing money by field double-counts. Any “which discipline gets the most” chart depends on whether you divided the grant across labels, assigned it to the first, or double-counted.
• status distinguishes signed, ongoing and closed projects. Filtering to closed is defensible and preferentially drops recent years.
• Nominal euros. No deflator anywhere. A 2014→2026 nominal series across the 2022 inflation spike overstates real growth substantially, and not deflating requires no code at all.
• endOfParticipation and active in organization.csv record participants who dropped out — including, notably, UK organisations around Brexit.

Sensitivity

Moderate — political, not personal. Nobody is individually harmed, but “the EU wastes money on research” and “country X is a net loser” are live political claims with existing constituencies. A polished poster arguing either, sourced to the Commission’s own data, is repeatable.

Keep claims about definitions and measurement rather than about desert: “EU research funding doubled in 2022” is a good jam claim because the refutation is technical and the topic is dry. “Country X freeloads on EU research money” is a bad one because the refutation is political and the poster will be screenshotted. F2 labelling matters here.