Official figures and this site's calculations

Method and sources

Where the figures come from and how they were computed, with the files to download; who makes the site, and the corrections made to it.

Method

Study verdicts and ranks are classifications made by this site; each cites the phrase it rests on. Other numbers are shown with their source, definition and year, and with the calculation from the source to the page. Where no source was found, the page says so.

ObservedA published accounting or administrative figure, used as published.
EstimatedA model or an imputation by the publisher, such as distributional accounts.
HypotheticalA scenario.

The figures are also in claims.json, each with the Eurostat cell or the document it was read from. Eurostat cells can be re-queried one by one from their panels.

About

Who makes this site

This site is made by Thomas Gazagnaire, as a personal project. It has no funding and no institutional affiliation.

Its code reads the sources and renders the pages, every number with its provenance: the Eurostat cell or the document it was read from, or the calculation that produced it. The code and the data are at github.com/samoht/budget, from which any page can be rebuilt; the footer of each page gives the date and the commit it was built from.

Corrections and replies

To report an error, open an issue at github.com/samoht/budget/issues. Since 16 October 2026, each correction is logged below, latest first, with the pages it changed, what changed and why.

An institution or an author named on the site may ask, through the same issues, for a reply to be published next to the passage concerned.

Corrections log

2026-10-17 · Method · Found in review
The ◐ mark is defined as an estimate, by its publisher or by this site's evaluations. A count this site makes of its own data says nothing more to a screen reader, and a figure computed from such counts says “calculated”.
The glossary defined ◐ as a publisher's model while this site's own estimates and counts carried it.
2026-10-17 · Home, Public money · Found by the author
The 2027 rows and the paragraph on the draft budgets say that the Assemblée nationale deleted the bill's article liminaire in its sitting of 14 October 2026.
The figures are those of the bill as tabled, and the article that set them out is no longer in the text.
2026-10-17 · Home, Does it work? · Found in review
The summaries state their rule: a result counts when its 95% interval excludes zero and it passes its design's checks, no gap before the change and a placebo that finds nothing. Two do, the DSR saving and the DETR's first-year jump against communes' own past; the other results that exclude zero are named with the check they fail or lack. The pages had said three results exclude zero, the fall in prices where Pinel closed among them.
Other intervals on the study pages exclude zero too. The price fall is the median of all sales, mostly existing dwellings, with one year before the change, and zone A, where no rule changed, shows a fall that cannot be told from it.
2026-10-17 · Home, Does it work? · Found in review
The share of public money in single measures with a causal study, 6.1%, is given with the total it divides, €1,845bn, and without the general cuts in employer contributions, 1.9%.
Those cuts hold €77.6bn of the €111.8bn, and their studies cover their introduction in the 1990s and the CICE's conversion in 2019, not the current scheme.
2026-10-17 · Does it work? · Found in review
The register section says that the line of research and development for the economy may hold the research tax credit, which has causal studies, and that the State's grants to local authorities weigh €0.1m because consolidation removes them.
Insee does not say which function holds the credit, and the grants this site evaluates sit on a line the consolidated accounts empty.
2026-10-17 · Home · Found in review
The home page's figures by tenth say that the richest tenth's receipts are mostly pensions, counted here as transfers.
Without pensions, the richest tenth received less than the poorest.
2026-10-17 · Home, Public money · Found in review
The home page no longer gives a euro figure for the top 0.04%; it gives the IPP's rates and points to the money page, which gives the range, names the assumption that overstates their taxes on products, and says the top of the range comes from the profits kept in companies.
The range does not bound that overstatement, and its top end measures attributed company profits, not taxes.
2026-10-17 · Public money · Found in review
The summary of where public money goes gives each line's share of all public spending, €1,672bn, and names that base: old age is 23.4%. It had given 44.8%, a share of social protection and general services only.
The summary sat beside a sentence giving shares of all spending.
2026-10-17 · Home, Public money · Found in review
Public social spending in the OECD comparison is named gross, and 2019, the year before the pandemic, is given beside 2021: net total social spending was 30.1% of GDP in France and 29.2% in the United States.
2021 was a year of pandemic spending.
2026-10-17 · Public money · Found in review
The State's spending norm is described as the 2023-2027 programming act sets it, one aggregate that leaves out debt interest and tax refunds; its 2026 figure is what the finance law for 2026 set, and the Ondam for 2027 is what the bill proposes.
A finance law sets the norm without voting it, and a bill's Ondam is a proposal until Parliament votes it.
2026-10-17 · Wealth · Found in review
The minimum tax on high incomes brought in €0.4bn in 2025, the instalment of December 2025 on 2025 incomes; the page had said it raised €400m for 2025. The extension clause is quoted from article 2 of the finance law for 2026.
The Cour des comptes writes that the final yield on 2025 incomes will be known only in July 2026.
2026-10-17 · Wealth · Found in review
This site's arithmetic of the 2% minimum tax, €17.3bn and €24.6bn, is said to be an upper bound at its base.
The proposal deducts the taxes already paid, and the arithmetic does not.
2026-10-17 · Home, Does it work? · Found in review
The DSR saving is dated, 2016-2019, scoped to the fraction cible, €324m in 2019, and read beside the rule that a commune balance its operating budget and borrow only to invest.
The summaries gave the saving without its years, its scope or the budget rule it may reflect.
2026-10-17 · ZRR, Political colour, Apprenticeships · Found in review
Three phrasings: the bunching of EPCIs under the ZRR's density threshold is stated without imputing an intent; close municipal elections are those won or lost within 10 points of the vote, not by a few votes; the hiring aid is said to have lapsed because the Government let it lapse until the finance law and its decree, and the 2026 fall under reduced aid is given beside the 2025 fall under equal aid.
The pages said more than the data show, or less than the design uses.
2026-10-17 · Method, Retirement age · Found in review
The glossary defines a synthetic control as built from units that did not get the same change, or got it later, and the pensions page says its donor countries were raising their own pension ages by other rules, so it compares France's reform with theirs.
The two contradicted each other.
2026-10-17 · Local grants · Found in review
The State's DETR indicator is set beside the same ratio of accounts recomputed from the DGCL lists, 4.96 in 2025; the causal leverage at the eligibility cap is beside it as another quantity, with its specification.
The causal leverage is not the quantity the State's indicator measures.
2026-10-17 · Home, Does it work?, Public money, Wealth, Local grants · Found in review
Every panel of a figure this site calculates says “Calculated by this site” beside its status.
The status alone did not tell a figure read from a source from one this site computed.
2026-10-17 · Home, Apprenticeships · Found in review
The apprenticeship pages say that since 1 July 2025 the employer of an apprentice preparing a niveau 6 diploma or above pays a flat €750 towards the contract's training costs, under décret n° 2025-585 du 27 juin 2025, which the months of 2023 and 2024 the comparison starts from did not have.
The comparisons of March-December 2025 and of early 2026 include that cost, which the pages had left out.
2026-10-17 · Does it work? · Found in review
The table of impact, leakage and permanence applies one rule to every study: a verdict counts only against a goal of the line, from a causal design; money that reaches people the law does not target, such as landlords under housing benefit, is capture, a leakage, recorded apart from who ends up with the money among those it targets; a study of a cut is marked; and the 23 findings that rest on a model's simulation are ringed and left out of the counts.
The studies had not all been coded by one rule: capture by landlords and sellers was filed with who ends up with the money, a study of a cut was not marked, and a simulation could make a goal measured or a line studied.
2026-10-17 · Does it work? · Found in review
On the French page, the tooltips of the evaluations table and the panels of its studies name their authors and this site's evaluations in French; on both pages, a study listed with "et al." among its authors reads "Payen A et al.", not "Payen A and et al.".
The names were joined in English when the data was read, before the page's language applied.
2026-10-17 · Companies · Found in review
The general cuts in employer contributions are shown at €77.6bn for 2024, all compulsory schemes, in the chart of the largest schemes and in the evaluation table, the figure the evaluations page and the home page use; these two had shown €64.9bn, the basic social-security schemes alone.
The two figures are the same reductions in two scopes, read from the same annex of the social security budget; the page now uses the wider one and states the narrower one beside it.
2026-10-17 · Companies · Found in review
For the CICE, the share of the credit passed into wages is recorded under who gets it, as a fact; the page had written it under leakage. The finding that the job effect is significant only for independent firms, which that cell held and which does not say who receives the credit, is no longer shown.
Who ends up with the money among those the law targets is recorded apart from leakage, as the evaluations page's rule says.
2026-10-16 · Home, Does it work?, Retirement age · Found in review
Since the 2023 pension reform, employment at 60-64 has risen more than at 55-59, so the gap between the two age bands narrowed; the pages said it grew.
Employment at 60-64 is lower than at 55-59; as it caught up, the difference between the two moved towards zero, which is a narrowing.
2026-10-16 · Home · Found in review
France beside Germany: old age, mainly pensions, accounts for 3.5 of the 7.6 points of GDP by which France's public spending exceeds Germany's, and social protection as a whole for 3.3; the page said social protection made up the larger part of the difference.
Computed from the Eurostat cells, social protection is less than half of the difference; each figure is now shown with its panel.
2026-10-16 · Home, Retirement age · Found in review
The pages said the gap between employment at 60-64 and at 55-59 also moved in 2005-2010 “without any reform”; they now say with no change in the legal minimum age, and the pensions page names the reforms of those years.
Those years saw no change in the minimum age, but the 2003 reform's early retirement for long careers and its décote and surcote were phasing in, the job-search exemption for older unemployed people was withdrawn from 2009, and retirement imposed by the employer before 70 ended in 2010.
2026-10-16 · Home · Found in review
The 2025 spending tile now says tax credits included, the 2026 tiles give the Government's own 2025 figures beside its forecasts (56.6% of GDP and €1,694bn of spending, 43.6% of GDP in levies), and the note says the two rows are not a change.
The 2025 row is Eurostat's, with tax credits counted as spending; the 2026 row is the Government's, without them. Read one after the other, the two rows showed a change that is a difference of definition.
2026-10-16 · Taxes · Found in review
The Dutreil exemption: the budget documents put its cost at €0.5bn up to the PLF 2025 and at €5bn in the PLF 2026; the page gave only the €0.5bn.
The PLF 2026 revised the estimate (fiche 520110), and the page had kept the earlier one.
2026-10-16 · Companies · Found in review
The CICE row of the evaluation table said the credit's 2019 conversion into a cut in employer contributions was never evaluated ex post; it now cites the IPP's ex post evaluation of that conversion (report no. 38, 2022), which detects no effect on employment, value added, turnover or investment in 2019 and 2020.
That evaluation exists, and the home page already quoted it.
2026-10-16 · Public money, Home · Found in review
Along the top tail of the distribution, every levy is spread in proportion to income; the page now says this is this site's assumption, not Insee's attribution keys, and that it overstates what the very top pays in VAT and production taxes. The home page gives both ends of the range.
Insee attributes taxes on products by consumption, and in its own twentieths taxes on products and on production rise more slowly than income.
2026-10-16 · Home · Found in review
The home page now carries each study page's reading of two falls: in private jobs in the rural tax zone, better read as no gain than as a loss caused by the zone; and in industrial jobs in the industrial territories, a census difference that no placebo can check.
The home page stated the two falls without the caveats of the pages they come from.
2026-10-16 · Home · Found in review
The home page's title said “France, in six questions” over five question pages; it now says five.
The sixth card, Method, is not a question.
2026-10-16 · Home, Does it work? · Found in review
Public money counts the payable tax credits once, inside spending, where the national accounts record them. The share of public money in single measures with a causal study moved from 6% to 6.1%, and the upper bound for broad lines from 82.2% to 82.3%.
The credits were counted twice, as spending and as tax expenditures.
2026-10-16 · Home, Does it work? · Found in review
The research tax credit's effect on researchers is now among the results these estimates cannot tell from zero; the summaries had counted it among three changes.
Its p-value, 0.10, is the smallest a ranking among ten countries allows.
2026-10-16 · Vacant dwellings, Does it work? · Found in review
The 292 communes added to the vacancy tax in 2024 are described as the communes the decree of 25 August 2023 added to the list of tight urban areas of more than 50,000 inhabitants; the page had described them as outside the large urban areas.
The decree redrew that list; the communes outside it are the 2,243 tourist communes of the 2023 finance law.
2026-10-16 · Vacant dwellings, Housing zones, Does it work? · Found in review
The vacancy and zoning pages now open on what each design can tell: vacancy moved by +0.13 per 100 dwellings against the communes taxed since 2013, whose rates rose in 2023, and the zoning estimates run from −1.00 to +0.35 flats per 1,000 inhabitants depending on the estimator. Both had opened on “no fall”.
Each lede rested on one comparison or one estimator, and the others did not support it.
2026-10-16 · Home, Does it work?, Apprenticeships · Found in review
The apprenticeship hiring aid for higher-education contracts is said to have lapsed in January and February 2025 and 2026 until the year's finance law and decree; the pages had said it was withdrawn.
The exceptional aid is set by decrees with an end date, and decrees resumed it once the finance law was passed.
2026-10-16 · Wealth · Found in review
The levy calculator now starts from the proposals tabled since 2025, each with its text, its authors' or an official costing and this site's arithmetic; the rate that would repay the debt and the yearly mode on a fixed base are gone.
Those settings were this site's, not those of any proposal tabled.
2026-10-16 · ZRR · Found in review
The rural tax zone's page no longer opens on the carbon-offset method with communes in place of forest pixels; the method it borrows is named in the body.
The image described where the method came from, not what the page measures.
2026-10-16 · Home, Does it work? · Found in review
The summaries recounted this site's evaluations as two descriptive changes and the results whose intervals exclude zero, each with its design; they had counted three changes.
The apprenticeship and pension changes are descriptive, and the research tax credit result cannot be told from zero.
2026-10-16 · Does it work? · Found in review
The register codes each study on the goal it measures, on three levels: measured, touched only indirectly, or not measured. Who ends up with the money, through wages, rents or parents, is recorded as a fact, apart from leakage.
A study of another outcome does not measure a line's goal, and money reaching the people a law targets is not leakage.
2026-10-16 · Home, Does it work? · Found in review
Of the 155 official goals, a causal study measures 50 and touches 10 more only indirectly; the pages had said 60 were measured.
Ten of the sixty were reached only through a neighbouring outcome.
2026-10-16 · Home · Found in review
Halving classes in priority-education primary schools is said to pay for itself in the long run if its early gains persist, on a calibrated projection, beside the studies that find modest effects that partly fade; the page had said it pays for itself, from causal estimates.
The claim rests on a projection, not on a causal estimate of long-run gains.
2026-10-16 · Home, Companies · Found in review
The Senate's broad count of support to companies, €211bn, is said to include €41bn of Bpifrance loans, guarantees and equity at face value, and its rapporteur's reason for calling it a floor is given: regional, local and EU aid are left out.
The page called the count a floor without saying what it includes at face value.
2026-10-16 · Home, Public money · Found in review
Lending, interest and debt are attributed to general government, not to the State.
The figures are those of general government, which includes local authorities and social security.
2026-10-16 · Local grants, Does it work? · Found in review
The DSR result is said to hold where the grant comes and goes from year to year, and the page says it cannot tell how steady recipients use the grant; the summaries had said the DSR is saved. The DETR design is said to have the power to detect only a leverage of at least about 12, not to detect one of about 12; and the State's objective for the investment grants is quoted in full.
Each sentence said more than the design or the source supports.
2026-10-16 · Housing zones · Found in review
The cost of the buy-to-let reductions, €1.5bn in 2024, is attributed to Pinel and Duflot together; the moves up the zoning in 2023-2025 are said to rest on provisional counts in the year or two after the move.
The cost line covers both schemes, and the counts for the latest years are provisional.
2026-10-16 · Retirement age · Found in review
The pension reform is said to be partly suspended, and the placebo is read as what the method finds in a period of other reforms; the page had said suspended, and called the placebo noise.
The suspension covers part of the reform, and the placebo years saw other reforms.

Glossary

The terms the pages use

The methods of the evaluations and the categories of the public accounts, in one or two sentences each.

p-value: The probability of a result at least as far from zero as the one found, if the scheme had no effect at all. With few units, as when France is ranked among ten countries, the smallest p-value possible is large: one in ten.
95% interval: The range of effects the data are compatible with at the usual threshold. An interval that contains zero means the design cannot tell the effect from none.
Minimum detectable effect: The smallest effect a design would find four times in five, given its units and how much they vary. A result that cannot be told from zero says little about effects smaller than this.
Regression discontinuity: Compares units just above and just below a threshold that decides who gets the money, such as a population or a rank, which are alike but for the money. It measures the effect at the threshold only.
Difference-in-differences and event study: Compares how an outcome changed where the money went with how it changed in similar units that did not get it, over the same years. An event study shows that difference year by year, before and after the money, so that a gap already there before shows.
Synthetic control and synthetic DiD: Builds the comparison for one unit that got the money or the change, such as France, as a weighted mix of units that did not get the same change, or got it later, and that tracked it beforehand. Synthetic difference-in-differences weights the years too, and compares changes rather than levels.
Matching and placebo: Matching pairs each unit that got the money with one that did not but looked alike beforehand. A placebo runs the same design where nothing happened, at a fake date or on units the scheme did not reach: an effect found there means the design picks up something other than the scheme.
Pre-trend test: Checks that the units that got the money and their comparison moved together before it arrived. If they were already drifting apart, the difference afterwards cannot be read as the scheme's effect.
COFOG: The United Nations classification of government spending by function, which Eurostat uses: ten divisions, such as social protection or education, split into groups, such as old age or primary education. The spending lines of the site are its groups.
Public money: What the evaluations page adds up for 2024, line by line: general government spending by COFOG group, plus the tax expenditures and the exemptions from social contributions that serve public aims through the tax system, with the payable tax credits counted once, inside spending. It is the denominator of every share of money with a causal study. Tax expenditures are each costed with the others unchanged, so the total is an order of magnitude.
Leakage: Whether the effect of the money moved elsewhere: activity displaced from places or firms the scheme does not reach, a windfall for those who would have acted anyway, or capture, money that reaches people the law does not target, such as landlords under housing benefit or sellers under a tax incentive. The evaluations page counts it as a verdict on the money.
Incidence (who ends up with it): Who ends up with the money among those the law targets, such as workers under a wage subsidy, tenants, or the parents of grant holders. The evaluations page records it as a fact, not as leakage: its square has no fill.
Touched: A goal a study reaches only indirectly, as sick-pay studies measure sick days, not the workers' income the law guarantees. The study's reading stays in its panel, draws a grey square and is not counted.
Simulation: A finding that rests on a structural or calibrated model's simulation, not on a measured comparison. The evaluations page keeps it, ringed, and leaves it out of the counts: a simulation alone does not make a goal measured or a line studied.
S13, general government: The sector of the national accounts that holds the State, local authorities, social security funds and the bodies they control. Public spending on the site is the spending of S13, as Eurostat publishes it.
Potentiel financier: What a commune's taxes would raise at the national average rates, plus the fixed part of its general operating grant, per inhabitant. The State uses it to decide which communes receive equalisation grants such as the DSR and the DETR.
Tax expenditure: A provision of tax law that lowers the tax some taxpayers pay in order to serve a public aim, costed each year in Voies et moyens, tome II. Each is costed with the others unchanged, so their sum is an order of magnitude.
Payable tax credit: A tax credit paid out in full even where it exceeds the tax due, such as the research tax credit. The national accounts record it as spending, so the site counts it once, inside spending.
Living standard per consumption unit: A household's disposable income divided by its consumption units: 1 for the first adult, 0.5 for each other person aged 14 or over, 0.3 for each child under 14. It is Insee's measure of living standard, and lets households of different sizes be compared.
“Calculated by this site”: Marks a finding that rests on a calculation or an evaluation of this site, rather than on a figure read from a source. Its panel shows the calculation.
◐ estimated and ○ hypothetical: A figure marked ◐ is an estimate: a model or an imputation by its publisher, such as distributional accounts, or an effect this site estimates in its evaluations; a screen reader says “estimated” after it. A count this site makes of its own data, such as the official goals a study measures or the communes a design pairs, also carries ◐ where its rows rest on the site's matching or classification; the count itself is exact, so a screen reader says nothing after it, and “calculated” after a figure computed from such counts. One marked ○ is a forecast or a plan, which a screen reader calls a forecast, or a scenario resting on a parameter this site chose, which it calls hypothetical. An unmarked figure is a published accounting or administrative figure, used as published.

Data

Get the data

The figures on the site are in claims.json and the time series in series.json. The results of the evaluations run by this site are in the CSV files listed below.

The evidence panels of the Public money page's diagram of who pays and who receives, in the page's language, split into parts of a few hundred kilobytes, each view and year listing its parts on one page: evidence/money-tenths-2023.html, evidence/money-tenths-2024.html, evidence/money-shares-2023.html, evidence/money-shares-2024.html, evidence/money-ages-2023.html, evidence/money-ages-2024.html, evidence/money-persons-2023.html, evidence/money-persons-2024.html.

The code that reads the sources and builds the pages is at github.com/samoht/budget.