Editorial standards

These are the rules every piece on this site is held to. They describe what we actually do, not what we aim for. Where a rule has been broken, the piece carries a dated correction, listed on the corrections page.

Sourcing

• Every number names a public source a reader can fetch again, and the piece links to it: a series, a dataset or a named document. A figure that cannot be traced to one is not published.
• No number is written from memory. Each figure is fetched or computed from its source for the piece, and recorded with its date, source and method before it appears in the text.
• Lag is stated. Official data arrives late. A figure is given with the date it describes (“the latest published, for 6 October”), never as if it were today’s.
• Records are checked by scanning the whole series, and are stated within the series’ coverage: “the lowest in PortWatch’s data, which starts in 2019”, not “the lowest ever”.
• Correlation is not reported as cause. When and how much are usually in the data; why usually is not. A mechanism is given only when a primary source states it, and is attributed to that source.
• Quotes are verbatim and linked. There are no anonymous “analysts say”, and no paraphrase is turned into a quote.

The data comes from the institutions that publish it, including:

• the Federal Reserve Bank of St Louis (FRED), the Federal Reserve Bank of New York, the US Bureau of Labor Statistics and the US Treasury;
• the European Central Bank, Eurostat and the UK’s Office for National Statistics;
• the IMF, including IMF PortWatch for shipping, the World Bank and the OECD;
• Germany’s waterways agency, WSV (PEGELONLINE), and the US National Weather Service for river levels.

The site’s own measures are built from these and published with their methods: the Stress Index, the chokepoint board and the Outlook.

Fact-checking

Before a piece is published, a separate check goes through it claim by claim. It is run fresh, without the writer’s working, so that it does not inherit the writer’s mistakes:

• every number is re-derived from the raw data, not read back from the draft;
• every date, record claim, quote and attribution is traced to its source;
• every link is opened;
• the headline and standfirst get the most scrutiny, because they are what most readers see.

A single failed claim stops the piece until it is fixed. The piece then passes a quality gate: house rules, a scored editorial review, and a build of the site with the piece in place. A piece runs only when the data has a finding the archive does not already hold, so a day with nothing published is a normal outcome.

Statements by public figures

Pieces in “On the record” report what a head of government, minister, central banker, regulator or large company’s chief executive said in public. Each also gives the other parties’ cases, in their own words:

• Quotes come from the primary record. That means the post itself, the transcript or the official release. When the original cannot be opened, the exact words must appear in two independent outlets, and the piece names them.
• Each party’s position comes only from that party’s own statements and actions. Where a party has not responded, the piece says so, with the time of the check.
• Hedges are kept. An “if”, an “unless” or a “for now” is never cut.
• No motive-reading. We report what was said and done, not why.
• The piece states its cutoff, the time up to which it reflects the public record. A later reply or deletion is a reason to update it, and the update is dated.

Forecasts

We do not make predictions in articles. The only forward-looking numbers on the site are on the Outlook page, and each has to earn its place:

• each model is tested the way it is used, on data that was public at the time;
• it must beat the simple no-change forecast;
• it is published with its 80% range and its record of how often outcomes landed inside that range.

Every forecast is logged when it is first issued and scored against the official figure when that appears. A model that stops clearing the bar is switched off automatically, and the page says why. Forecasts that are not ours, such as the agencies’ river forecasts and the New York Fed’s recession model, are labelled as theirs.

AI assistance and human review

Pieces are researched, drafted and checked with AI assistance (Claude, by Anthropic). The rules above are enforced at every stage, including scripts that refuse to publish a piece that fails its checks. A person reviews every piece and approves its publication, and every piece is labelled “AI-assisted · human-reviewed”. If that review were ever skipped, the label would not be used.

Corrections

We correct errors openly and promptly:

• Never silently. A correction changes the text and adds a dated note at the foot of the piece, saying what was wrong and what it says now.
• Listed in one place. Every correction is on the corrections page, newest first.
• Revised data is not an error. Publishers revise their data. When a source revises a series after we publish, the piece stays as written for its date. If the revision changes a finding, we add a dated update.
• Out-of-date facts are updated. The topic guides and explainers are kept current. When a figure in them was simply out of date, we update it. When it was wrong, it gets a correction note like any other piece.

To report an error, write to [email protected] with the page and the claim. We reply, and if we agree, we correct it.

Independence

Our analysis is not influenced by affiliates, advertisers or data providers. Data providers supply raw data. They have no editorial input, no review rights and no influence over what we publish.

Contact