Sources & methodology

The three sources

Nothing is read from a mirror, a scraped dashboard or a reconstructed dataset. Each source is queried at its own official endpoint, and the request parameters are stored beside the snapshot.

The Combined Nomenclature is established by Council Regulation (EEC) No 2658/87 and re-adopted each year. Where a legal question turns on exact wording, the Official Journal is the authority — not this snapshot.

Codes: HS4 and CN8 are not the same thing

A customs code gets longer as it gets more specific. The first six digits are the Harmonized System, shared worldwide. The European Union adds two more, giving the eight-digit CN8 line an exporter actually declares.

• HS heading 1902 (HS4) covers pasta, couscous and similar preparations as a whole — every kind of pasta, stuffed or not, plus couscous.
• 1902 19 10 (CN8) is one line inside it: uncooked pasta, not stuffed, containing no eggs, of durum wheat only.

Requested code vs observed code

The code you ask about and the code the data was collected at are tracked separately, and their relationship is stated rather than assumed. This is what customs coverage means in the interface.

Who reports what, and at which value

A shipment is counted twice in world statistics: once by the country that sent it and once by the country that received it. The two figures are never equal, and they are never added together here.

• Import reporter. The destination declares what came in. This is the basis of Market Intelligence: it is the buyer’s own count of its market.
• Export reporter. Italy declares what went out. This is the basis of the Eurostat view, and it answers a different question.
• CIF and FOB. Imports are valued CIF — including cost, insurance and freight to the border. Exports are valued FOB, at the point of departure. The same shipment is therefore worth more as an import than as an export, by roughly the cost of getting it there.
• Currency. UN Comtrade reports in current US dollars; Eurostat reports in euro. The two are never combined into one figure, and no exchange rate is applied to make them look comparable.

WORLD is not the sum of the partners

Statistical agencies publish a “world” total alongside the per-partner rows, and the two do not match exactly: confidential flows, low-value thresholds and unallocated trade land in one and not the other. Where a partner breakdown is shown, it is the sum of the partners and is labelled as such. In the Eurostat snapshot the partner rows come to 99.99% of the published world total for 2023 — close, but not the same number, which is why they are never used interchangeably.

Observed, derived, estimated, company input

Every figure in the product carries one of four labels. They are not decoration: they are the difference between a fact and an assumption, and the interface prints them beside the number rather than in a legal note at the bottom.

Years, and why 2024 is marked partial

Countries file at different times, so the most recent year is always incomplete. In this snapshot 2023 has 164 reporters and 2024 has fewer — the large reporters have filed and many smaller ones have not.

The awkward part is that 2024’s total is nonetheless higher than 2023’s. Comparing the two would read incomplete reporting as growth. So 2023 is the common reference year, 2024 is kept and marked partial everywhere it appears, and no growth figure is ever computed across it — the year-on-year field returns “not comparable” with the reason attached rather than a silent blank.

Supplier detail exists for 30 markets

Knowing who else supplies a market requires the full partner breakdown for that market. The public data tier returns 500 records per call and one year per call, so the whole bilateral matrix is not obtainable. It was fetched for the 30 largest importing markets, chosen by computing them from the demand data rather than picking them by hand.

Those 30 markets account for the large majority of observed demand. Every other market reports supplier detail as unavailable — with the reason — rather than having it approximated from its totals. Concentration measures are withheld entirely when the partner rows do not form a complete partition of the market, because a concentration figure computed on a partial breakdown is simply wrong.

The supplier matrix covers a declared subset of markets, not the world. It is never described as global coverage.

The company figures are a demonstration

Company Scenario runs on a fictitious 500 g pack of dry pasta across 8 demonstration markets. The customs code is official; the price, the five unit costs, the freight, the tariff rates, the exchange-rate exposure and the demand response are all demonstration values, chosen so the engine can be seen working.

Anything you type stays in your browser. The shareable link is assembled in the browser from the numbers and read back in the browser, and it deliberately omits the product name — the one field that could identify a real SKU.

Scenarios, not forecasts

The engine propagates the assumptions you chose through explicit arithmetic. It does not say what will happen, and no probability is attached to any result. Move the raw material lever and you learn what a 25% input rise would do under the stated assumptions — not how likely that rise is.

The levers currently modelled are external shocks: things that happen to you — input prices, oil, exchange rates, tariffs, and how much of it you pass on in price. Company interventions — trade fairs, campaigns, influencer work, listings — are named in the interface and deliberately not modelled. The engine moves demand only through price, so pushing a marketing budget through a price elasticity would produce a confident-looking number that stands for nothing.

There is no opportunity score

A single number ranking markets would be the easiest thing here to build and the least honest. It hides a weighting — how much demand size counts against growth, against share, against competition — inside a figure nobody can inspect or argue with.

Instead there are 7 questions. Each is one observed measure with one stated inclusion rule and one ordering. Nothing is weighted against anything else, nothing is combined, and no market is ever labelled best, recommended or promising. Two questions can order the same markets in opposite directions, and that is the intended behaviour: you are choosing what to look at, not being told what to do.

One judgement is made and declared: a data-sufficiency floor. A share or a growth rate computed on a market that imported forty thousand dollars is arithmetically fine and practically noise, so views built on ratios exclude markets under $10m and say so on screen. Size itself is never floored.

How methods are governed

The refusals on this page are not editorial preferences. They are enforced by a written methodology that the code is checked against, so that neither a future change nor an automated agent can quietly loosen them.

• Versioned models. Every model a figure comes from has a dated specification stating its objective, its unit of analysis, its formulas, its gates, what it is allowed to claim and what it is forbidden to claim.
• Registered indicators. Every indicator is recorded with its formula, its unit, its source, its direction — or the fact that it has none — and its rule for missing data. A missing value is never a zero.
• Traceable sources. Every methodological claim cites a primary or official source, recorded with the date it was consulted and, as importantly, a note on what that source does not support.
• Declared heuristics. Where this product chose a band or a floor rather than deriving it, it is labelled as this product’s choice. No threshold here comes from economic theory or from measurement, and none is presented as though it did.
• Sensitivity. Where a choice changes which markets you see, the product computes the answer one step either side of it and shows you which conclusions survived.
• Prohibited claims. Each model carries an explicit list of things it may not be used to say — no composite score, no forecast, and no recommendation aimed at a single company, because a national figure is not a company outcome.
• Review status. Every model has been through one independent methodological review. None has had a second, so all are marked provisional rather than approved, and the review’s open findings are recorded rather than closed.

The full framework, the indicator registry, the model specifications and the review findings live with the code in the repository’s docs/methodology/ directory, alongside the verifier that checks the code against them.

What the pilot does not cover

• One heading. HS heading 1902 only. Not every food and beverage product, and not the rest of the nomenclature’s agri-food scope.
• No CN8 trade figures. World trade is observed at four digits. See Codes.
• No forecast. No projection, no trend extrapolation, no probability.
• No recommendation. No score, no ranking of best markets, no suggested action.
• No company interventions. Fairs, campaigns and promotion are not modelled.
• No live data. These are snapshots with recorded acquisition dates, not a feed. A figure is as current as the date at the top of this page.
• No generative AI. Every sentence the product generates about a result is assembled by explicit rules from the numbers behind it.
• Demonstration company economics. See the company figures.

A deeper technical record — endpoints, request parameters, checksums, the refusals each acquisition makes — is kept with the code in the repository’s docs/ directory, alongside the verification scripts that pin the figures quoted here.

Start a decision session · Open Market Intelligence · Run a company scenario