From the source to the chart.

Every day an automated run visits each source: a government or international agency’s API, a published spreadsheet, CSV or PDF, or a port’s statistics page. The runs are spread across the day, so no publisher gets all our requests at once. Each dataset has its own collector, written for that one source, and each runs separately, so a source that is down holds up nothing else. Trade by partner and port (U.S. Census Bureau) and the country figures (IMF, World Bank, OECD, BIS and FRED) have collectors of their own, described on the trade methodology and each country page.

The Port of Long Beach blocks automated requests, so its monthly totals are typed in by hand from the port’s releases; the run reminds us when a month is due. A series can also be uploaded by hand, and keeps its figures until the next upload.

Collectors stop on anything unexpected: a page in place of data, a changed layout, parts that don’t add up to their total, or disagreement with a second copy of the same figures where one exists.

Before anything is published, a new download is held for a person to review if it has markedly fewer observations than the site already shows, ends earlier, or has an observation with no value. For port container counts, a past month that moves by more than 0.5%, or a new month that doesn’t follow the last one, is also held. A held or failed dataset keeps its published figures, and we are emailed.

Figures are used as the source publishes them wherever possible. Each dataset’s methodology lists anything we change: a unit (metres to millimetres, dollars to millions), how periods are dated, a sum of parts (the San Pedro Bay total is Los Angeles plus Long Beach, month by month) or a rebuilt aggregate.

We don’t seasonally adjust, smooth or fill gaps in the series we republish. A month the source didn’t publish stays empty, and the dataset’s methodology says so. The chart’s other views (changes, averages, volatility, constant dollars) are worked out from the published figures as you choose them, each with a note on how.

Two exceptions, both labelled where they appear. The CK Momentum Index, our own indicator, averages its inputs over three months, takes the usual seasonal pattern out of the port counts and fills October 2025, which the federal shutdown left unpublished, from the months either side. The constant-dollar view fills that same month of the price index by the same rule, for deflating only.

Constant dollars: each value × the base year’s average price index ÷ the index for its own month (for annual series, the average for its own calendar or federal fiscal year). A year’s average is the mean of its published months, as BLS calculates it, and counts once its last month is out with at most one month missing. A single month BLS didn’t publish (October 2025) is deflated by the geometric mean of the months either side; the price index datasets themselves are never filled. Series wholly inside the Los Angeles area use the Los Angeles consumer price index; all others the US city average.

Many series carry a CK forecast, a statistical projection with an 80% range, re-estimated whenever the data change. The model depends on the series: exponential smoothing on the logarithm with a fading trend for container volumes and prices; a no-trend version for rates and cyclical series; a mean-reverting autoregression for indices such as supply-chain pressure; and a recent median for ship queues.

Each projection is backtested before it is shown. We refit the model at up to 40 past dates (one a quarter over the last ten years, or one a year for annual series) and score its forecasts against what then happened and against two simple benchmarks: no change from the latest figure, and the same period a year earlier. A projection runs only as far ahead (three months to at most five years) as its errors stay at least 5% smaller than the better benchmark’s and its 80% range holds between 65% and 92% of outcomes it wasn’t built from; a series that fails at the shortest horizon gets none. The 80% range is the spread of the model’s own errors in those backtests, so it shows how far off it has really been. Each dataset lists its horizon and backtest scores in “Sources & methods” under the chart.

Projections extend past patterns and are labelled “CK forecast”: statistical projections, conditional on those patterns, not predictions of policy or shocks, and not the source’s own. Where a source publishes its own projections, we show those instead. The yearly forecasts for countries’ growth and inflation are a separate model, set out on the forecasts pages.