Five ways a customs table hands you the opposite answer

A UN Comtrade query for Peru's 2024 imports of HS 8705.40 from China returns twelve rows, and eleven of them are pieces of the twelfth. Each row carries three breakdown keys — customs procedure, mode of transport, and second partner — and the interface returns the total alongside every partial view of it. Add the value column down all twelve rows and you get USD 77,132,667 across 1,124 units. Take only the row where all three keys sit at their totals value (customs code C00, mode of transport 0, second partner 0) and you get USD 19,283,167 across 281 units. The ratio is exactly 4.000 on value and exactly 4.000 on units, because the same shipments are reported four times over: once as the total, once split by second partner, once split by transport mode, and once split by both. Nothing in the response marks the duplicates as duplicates.

The error is catchable without knowing the filter rule. Peru declared 386 units of this heading imported from all origins in 2024. A from-China figure of 1,124 units is 2.9 times Peru's own total from every country on earth, which is impossible rather than merely surprising. The corrected 281 sits below that 386 parent, and China's own export declaration for the same lane states 315 units — the ordinary size of gap between two customs authorities describing the same shipments. So whenever anyone shows you a country-level trade figure, ask what the containing total is. A child that exceeds its parent is an arithmetic error, and catching it needs no access to the source data at all. Row counts give no warning either: the identical query for Peru one year earlier, in 2023, returns a single row.

Twelve importing countries have a 2024 total row for HS 8705.40 in this dataset. Seven declared a quantity. The other five did not, and UN Comtrade records that in a flag most readers never open: isQtyEstimated, true for Saudi Arabia, Indonesia, Algeria, Nigeria and Thailand. Nothing else in the row marks the difference. An estimated figure renders identically to a declared one — same column, same formatting, no asterisk that survives an export into a spreadsheet — so a table sorted by unit count quietly mixes counted machines with modelled ones. The distinction is larger than it looks. A declared quantity is an observation about that country's trade; an estimated quantity is an observation about the dataset's own assumptions, applied to that country's declared value. One of those can support a market-size claim and the other cannot, and by the time the figure reaches a slide the flag that separated them is gone.

Two fingerprints identify them without reading the flag at all. First, the estimates are not whole numbers: 1,499.751 units for Saudi Arabia, 603.472 for Indonesia, 411.971 for Algeria, 285.93 for Nigeria, 54.785 for Thailand. No customs authority clears 603.472 machines, and every declared quantity in the same table is an integer. Second, divide value by quantity and the estimates collapse onto shared constants. Indonesia, Algeria and Thailand all return USD 58,536 per unit — 58,536.07, 58,536.06 and 58,536.24, the residual explained by the quantities being rounded to three decimal places. Saudi Arabia and Nigeria both return USD 63,315. Three unrelated economies do not import at the same unit price to five significant figures. They share a divisor, and the divisor is what the quantity was made out of.

The declared rows share nothing of the sort. Six of the seven carry a non-zero quantity, and excluding Malaysia, whose unit count is shown below not to be counting machines, the five remaining declared rows imply USD 50,852 to USD 82,160 per unit — a 1.6-times spread inside one heading in one year, which is what genuine national variation in a broad heading looks like. The seventh, Kazakhstan, declared USD 16,480,762 of imports against a quantity of zero: a third state again, neither counted nor estimated, simply absent. The operating rule follows from the mechanism rather than from caution. You may not derive a unit price from an estimated quantity, because a unit price was the input that produced it, and the answer you get back is the assumption you put in. The value column is declared in all twelve cases and stays usable throughout.

We asked fourteen exporting countries for their 2024 declarations of HS 8705.40. Ten returned a totals row. Four returned none: India, France, the United States and Mexico. The queries themselves succeeded and the dataset held nothing for those four, which is a different fact from a query that failed — and both render as an empty cell. Within the ten, China declared USD 800,257,504 across 18,017 units, against a ten-country total of USD 1,061,305,847. That is 75.4 per cent of the ten, and only of the ten. Written as a world share it is an upper bound, because every absent exporter can only enlarge the denominator. The honest sentence names the sample: 75.4 per cent of the declared value across the ten queried countries that reported for 2024.

The same table shows why the Peru double-count is not exotic. The number of rows a country's query returns tells you nothing about whether you are holding a total. China's returned one row, Italy's one, Japan's one. Germany's returned twenty-two, Spain's forty-one, Turkey's eighty-two. Germany's correct 2024 export total is USD 156,257,120 across 1,527 units, and it lives in exactly one of those twenty-two rows; take the first row the interface hands you and you are quoting a fragment with nothing on it saying so. Spain is a third variant: USD 10,779,213 of declared exports against a quantity of zero, a country reporting value while never reporting units. A share table that reads Spain's zero as zero machines is wrong in the opposite direction from Peru — understating rather than inflating, and just as invisibly.

China declared USD 122,709,885 and 2,383 units of HS 8705.40 exported to Russia in 2024. The twelve monthly rows sum to exactly the same figures, so nothing is lost in the annual aggregation except the shape. January to September accounts for 2,332 of those 2,383 units. October, November and December together account for 51. September alone was 686 units, more than a quarter of the year; November was 3. The nine-month pace is 259 units a month and the fourth-quarter pace is 17, a factor of fifteen. Annualise the fourth quarter and the lane runs at 204 units a year against the 2,383 the annual row reports. Both are correct arithmetic on the same rows, and they describe the same twelve months nearly a factor of twelve apart.

Which of those two numbers describes the lane a buyer or a seller actually met in 2025 depends on why the fourth quarter collapsed, and a customs table does not say. What it does say is that an annual total is the wrong instrument for detecting a closure, because a market that stops in October still reports three good quarters and still ranks where those quarters put it. Any country ranking built on annual figures alone will therefore keep a closed lane above an opening one for a full year afterwards, and the error corrects itself only when the following year's annual file publishes. The monthly series is the cheapest available correction, and it costs one more query against the same public endpoint.

The same series carries a second warning, and it is the one that produces confident errors rather than merely wrong ones. Every 2025 month returned no row at all. That is the monthly preview dataset not yet holding those periods; it is not a report of zero trade. A spreadsheet renders both as an empty cell and a chart renders both as a line falling to the floor, so an absence of data and a collapse of demand arrive looking identical and reading as a finding. When a series drops to zero in its most recent period, the first question is whether the data ends there or the trade does. The second is what the tool does when a query fails, because a failure that returns an empty result rather than an error will travel onward as a fact.

Malaysia declared 16,507 units of HS 8705.40 imported in 2024, worth USD 23,858,886. The quantity carries no estimation flag; it is what was declared. Divide it out and each unit is worth USD 1,445 and, against a net weight of 5,158,816 kg that Comtrade itself flags as estimated, weighs 313 kg. A concrete mixer truck neither weighs 313 kg nor costs USD 1,445, so whatever was counted here, it was not one machine per unit. On unit count alone Malaysia leads that twelve-country table outright, ahead of the Philippines at 2,463 units and Saudi Arabia's estimated 1,500. By declared value it is seventh of the twelve. A market ranking built on the quantity column would send a sales team to the country the value column ranks in the lower half.

The mirror settles it. China declared 403 units exported to Malaysia in 2024, worth USD 16,661,131, at a net weight of 5,517,832 kg. The two unit counts differ by a factor of 41. The two net weights differ by 6.5 per cent. The two sides agree almost exactly on how much steel crossed the border and disagree entirely on what one unit is, which is the signature of a counting convention rather than of a trade flow either side has mis-stated. For scale, every destination to which China declared 100 units or more of this heading in 2024 — thirty-two of them — falls between 11,623 and 16,978 kg per unit, with the all-destination average at 13,623 kg. So 313 kg is not a lighter machine. It is a different denominator.

The test is cheap and it belongs before any other analysis rather than after a surprising result. Divide net weight by quantity and ask whether the answer weighs like the machine. If it does not, the quantity column is measuring something else, and every per-unit figure derived from it is void regardless of how carefully the rest was computed. This case also shows how to rank your own evidence when the inputs are uneven. Malaysia's net weight itself carries the estimated-net-weight flag, so the weight ratio is corroborating rather than decisive; the unit value, declared on both sides of the mirror, is what makes the finding safe to state. Naming which half of the evidence is load-bearing is part of the check, not a hedge attached to the end of it.

All five failures have the same shape. The number is real and the sentence attached to it is not, so the questions worth asking are about the sentence rather than the arithmetic. Which side declared it, the exporter or the importer? One trade lane has two official values and they are not meant to match, because freight and insurance sit inside one and outside the other. Was the quantity declared or estimated? If estimated, nothing per-unit derived from it carries information about that country. Which rows were summed to produce it? If the answer is all of them, the figure may be a clean multiple of the truth, and the containing total will expose it without any access to the underlying file.

Then two more. What is the universe — which reporting countries are actually in the denominator, and which are missing from it? And what does the monthly series look like, because a year that ended in September still reports as a full year. None of these questions need paid data or a subscription. The UN Comtrade preview interface is public, the estimation flags travel inside the response, and each recomputation here took one query and one division. A supplier who cannot answer the five is not necessarily wrong; most people quoting trade statistics have never been asked. A supplier who treats the questions as hostile has told you which number to distrust, which is the cheapest thing you will learn in that conversation.