Every flow is the importer’s report where the importer filed, else the exporter’s

Every number on this site comes from UN Comtrade, the United Nations’ database of customs records, in its Final data set, standardised by the UN Statistics Division. Goods are classified in the Harmonized System (HS) and, before HS, in the Standard International Trade Classification (SITC). The annual data run from 1962 to 2025, the monthly data from Jan 2006 to Sep 2026.

A record says that one economy, the reporter, exported a product (a code) to one partner, or imported it from one, in a period (a year or a month), with a value in US dollars and often a net weight in kilograms. Exports are valued FOB, at the exporter’s border; imports CIF, with cost, insurance and freight to the importer’s border, so the same shipment is worth a few per cent more in the importer’s books. Imports are booked by country of origin, exports by country of last known destination.

The records are folded into a network layer on the machine that holds the Comtrade store. Reporters and partners become economies by ISO3 code (273 in all, historical areas such as the USSR and East Germany among them); groups, “areas n.e.s.”, bunkers and free zones drop out of the flows. The annual layer was built on 21 Oct 2026, the monthly layer on 23 Oct 2026, and the files this site reads on 2 Nov 2026.

The site carries 211 products: all goods, the 96 HS chapters, 29 baskets of codes that events turn on (chips, crude oil, vaccines and others) and the 85 stages of the ten chains. Products start in 1995, when HS was in full use; all goods go back to 1962 through SITC. Monthly data start in 2006 for all goods and in 2014 for baskets and chain stages.

One economy is read from a second source where UN Comtrade lacks its months. Korea files its monthly records with UN Comtrade once a year, four to five months after the year ends, and Comtrade’s monthly files hold Korea only from January 2013. For Korea, Jan 2000–Dec 2012 and Jan–Sep 2026, the site reads the Korea Customs Service, trade by item and country (data.go.kr) (source, fetched 31 Oct 2026). The records are summed to the same codes, in the HS edition Korea used at the time, and folded into the same partners as Korea’s Comtrade file; on the months both hold, partner totals agree to within 0.1%. The customs service’s item statistics leave out a few cells that Korea’s Comtrade file carries, about 0.04% of trade, among them Korea’s imports of lithium carbonate from Argentina. The last months, Aug–Sep 2026, are preliminary: they are drawn dashed and cited as such. The customs service revises its latest months, so they are fetched again every month after its mid-month release; each fetch is dated, revisions between fetches are logged, and when Comtrade’s file for a year arrives it replaces those months and the gap between the two is published with the checks. The customs records leave out Korea’s phosphate rock (2510.10) entirely, and a few other cells, about 0.04% of trade, among them its imports of lithium carbonate from Argentina; the stages concerned say so. Every chart and table that uses these months names the source.

Every flow between two reporters is in the data twice: in the exporter’s file and in the importer’s. They rarely agree. A CIF margin of a few per cent is expected (the IMF’s Direction of Trade Statistics uses a factor of 1.06); larger gaps come from re-exports and transit, with goods booked by origin on one side and by consignment on the other (Hong Kong, Singapore, the Netherlands, Belgium), from timing, confidential lines and misclassification. In 2017 at HS6 only 46% of records found their mirror; 32% were imports with no matching export and 22% exports with no matching import (Mitikj and Kaushik 2024).

Mirror rule. Annual views take the importer’s report where the importer reported that year (in HS, for products) and the exporter’s report otherwise, after Feenstra et al. (2005): importers report more completely and value goods at the border they cross. In 2024, 94.7% of world trade in all goods came from importers’ own reports.

Importer-reported and exporter-reported. The explorer can draw either side alone (m=importer, m=exporter). A flow card shows both reports and their ratio, the mirror ratio (importer’s report ÷ exporter’s); a ratio outside 0.8–1.25 gets a note.

Monthly views are single-sided. Monthly mirror data are noisy: US-reported imports of Chinese chips were 0.75 of China’s reported exports over 2023–24 on average, but the monthly ratio swung from 0.33 to 1.04. Stitching the two sides month by month would mix two timings, so each monthly flow takes one side over its window: the importer’s report when the importer filed at least as many months of the window as the exporter, else the exporter’s.

Non-reporters are seen only through their partners. Taiwan files as “Other Asia, n.e.s.” (code 490) and is shown as Taiwan. The EU and ASEAN file as reporters too, but they are groups and stay out of the network, so nothing counts twice. Historical areas (the USSR, Czechoslovakia, Yugoslavia, East Germany) are economies of their own.

Not every economy files, and many file a year or more late, so the latest periods look smaller than they are. 169 economies filed an annual file for 2023, 149 for 2024 and so far 123 for 2025.

Complete has one definition: the last period in which the ten largest traders all reported and coverage reached 98%. That is 2024 for years and Dec 2024 for months, when China’s monthly files end. Later periods are partial: their views carry a band and the share filed.

Coverage is the share of 2024 world exports (each economy’s exports to the world) held by the economies that filed the period; for months, of the economies that file monthly.

Missing is not zero. A pair absent from both reports may have traded unrecorded. When a large reporter is missing, its imports fall back to its partners’ export reports, which can look like a change in the network. Economies revise their files for years, so cite the build date.

Unit errors. A few files are off by orders of magnitude: Zimbabwe’s June 2006 monthly file reports $18 trillion of imports. The layer sets aside any file whose world total is more than ten times the median of the reporter’s neighbouring periods, and its partners’ reports stand in for it.

Incomplete periods. A recent period can hold an economy only in part: its own report is missing and some of its partners have not filed either, so their flows with it drop out of the period, and of every window that contains it. The site measures, for each economy and period, the share of its usual trade that is present: the value a year earlier of the flows still there, over the value of all its flows a year earlier (if that year was itself thin, the year before). A fall in trade does not lower the share; a missing partner does. Under 90%, the economy’s line stops, its card says how much is there and the table marks the period. The period picker ends at the last month in which at least 85% of world trade is present (Jul 2026).

The HS is revised every five years or so: H0 (1988/92), H1 (1996), H2 (2002), H3 (2007), H4 (2012), H5 (2017) and H6 (2022). Each reporter files a year in the version it used then, so in one year the same six digits can mean different goods in different economies’ files: in 2017, 60 reporters still filed in H4 and 106 in H5.

Codes are taken in HS2022. Where a code was split in 2022, the stage carries its older code too: solar modules are 8541.43 in HS2022 and sat with cells and LEDs in 8541.40 before; smartphones left 8517.12. Chapters and headings change less than six-digit codes, and long six-digit series need a concordance: Pierce and Schott (2012) found that revised codes covered 59% of US import value by 2004.

All goods go back to 1962 through SITC files; HS products start in 1995. Tier 0 of the checks looks for jumps in world trade at the 2017 and 2022 breaks, and flags a product when the jump is more than three times its usual yearly change and above 15%. In the last run, 84 products passed and 31 were flagged.

Monthly is the working frequency for anything with a date in it; annual is the frame. Seasonality is large, so it is named, not adjusted away: as an index of each year’s monthly mean over 2017–19 and 2022–24, China’s chip exports (8542, its own reports) run at 79 in February and 113 in September. We do not seasonally adjust bilateral flows, because thousands of short series with HS breaks would give factors no reader can reproduce. Every monthly chart says instead which sum it draws, in its caption and in the address.

Two explorers. The explorer draws years only (decision D6, 2 November 2026): the mirror rule, every network measure and every product back to 1995, all goods to 1962. Monthly draws months as rolling sums of 3, 6 or 12 months, with the months an importer has not filed estimated and flagged; it is the one place on the site to play the network month by month. Neither page draws a single month.

China’s chip exports in February 2023, against a year earlier, on this site’s monthly layer (each flow the importer’s report where the importer filed): the three months to February read −18% and the twelve months to February +1%. The same month is a fall in one reading and flat in the other; the reading is the story, so the caption always says which one is on screen.

January and February together. The Lunar New Year moves between January and February. China’s chip exports were $8.6bn in January and $9.7bn in February 2023 (holiday on 22 January) but $12.3bn and $10.1bn in 2024 (holiday on 10 February): single months read +43% and +5%, while the two months together grew 23%. Since 31 October 2026 no explorer draws a single month (decision D14), and no single-month value of a flow or an economy is published; the case studies’ partner shares keep January and February as one period, stored at February.

Base effects are shown, not hidden: a month that fell sharply makes the same month a year later read high. Strips offer “against two years earlier” for 2021–22 and “against 2019” for the pandemic.

Prices and the dollar. Values are current US dollars, so a price swing or a weaker euro or yen moves them. Where net weight covers at least 80% of value, a stage carries a unit value ($/kg) and a volume series; a change is called price-driven when the unit value explains more than half of it.

Explore (monthly) draws the network month by month on rolling sums of 3, 6 or 12 months (w), precomputed on the server. A window shows the turn of a series about (w − 1) / 2 months late: 1 month at 3, 2.5 at 6, 5.5 at 12. No window splits January from February, so no window ends in January, a 3-month window does not end in April and a 6-month window does not end in July. The 12-month sum is the default and equals lens L. Shorter windows move with the season, so their levels carry a warning and their change colours compare with the same window a year earlier (cmp=yoy). PageRank, betweenness and communities are computed on 12-month sums only; on 3 and 6 months the map shows sizes, flows and change colours.

Gap filling, after IMF DOTS. A flow is drawn from the importer’s report. Where the importer did not file a month and the exporter did, the month is estimated by the year-on-year splice of the IMF’s Direction of Trade Statistics (Marini, Dippelsman and Stanger 2018): est(m) = κ(m) × partner(m), where κ(m) is own ÷ partner over the 12 months ending a year earlier (both sides filed every month). Over a window this is own(t − 12) × partner(t) ÷ partner(t − 12) with the anchor taken on 12 months, which absorbs both the level gap between the two reports (CIF and FOB, re-exports) and the season. D14’s form anchors on the window itself; it gives a month a different weight in each window, and the published windows would then pin single months, so the page anchors month by month. The backtest scores both on the checks page: for all goods their median errors differ by less than a percentage point.

If the 12 months a year earlier were not fully filed, the latest fully filed 12 months within 24 months anchor the ratio. Beyond 24 months, or where a spliced month would exceed five times the importer’s own average month in the anchor year (a partner report that was near zero then), the month is the exporter’s report as is, flagged partner-only. Economies that never file monthly get the exporter’s report times 1.06 for CIF/FOB, also partner-only; reporting economies get no CIF/FOB factor. Once the importer’s annual file covers a year that holds estimates, only the estimated months are moved so that the year sums to the annual figure (benchmarked, proportional Denton; we found no maintained Python implementation of Cholette–Dagum, which Statistics Canada maintains in R as G-Series). Reported months are never changed.

Every flow and economy carries its estimate code, reported, benchmarked, spliced or partner-only. Estimated flows are drawn dashed and can be hidden (est=0); the caption gives the share of world trade that is estimated in the window on screen.

Divergence flags: detect, never correct. For each pair, the 12-month mirror ratio (importer’s report ÷ exporter’s, both filed every month) is flagged when it lies outside the median ± 3 MAD of the 36 months before, the band at least ±10% of the median wide: for all goods, and for chain stages at their HS2 chapters. The flag shows in the table and the tooltip, not on the map, and links to an event page when the event’s product and an economy of the pair match. It never changes a value; it says that the relationship between the two reports has broken. US imports of chips (8542) from China are the case in point: the yearly ratio (US report ÷ China’s) was 1.14 in 2022, 0.79 in 2023 and 0.66 in 2024.

Licence. Like every file on the site, the monthly explorer stores derived records only: shares of world trade, ranks, ratios, estimate codes and one world total per window, with no original record. No single-month value of a flow or an economy is stored or can be computed: a flow enters a window only when it is positive in every month of it, the mirror ratio only when both reports are, and the build tests the largest flows’ windows for a single month in their span and refuses to finish if it finds one.

What the page cannot do. Timing noise at product level remains: a 3-month window removes only part of it (see the timing share on the checks page), and chain stages swing more than all goods. Estimates are not reported data; they say what the importer would probably have reported, and they move once it files. The flags find a broken relationship; they do not say which report is right.

Source of the method: Marini, M., R. Dippelsman and M. Stanger (2018), “New Estimates for Direction of Trade Statistics”, IMF Working Paper WP/18/16. Decision D14 of the build plan.

The Harmonized System sorts goods by material and, within a material, by how far it has been worked, so within one material the code order is already a ladder: copper concentrate (2603.00), refined cathode (7403.11), wire (7408), cable (8544). Across materials the code list says nothing: no HS rule says that lithium-ion cells (8507.60) are made from cathode powder (2841.90) and graphite (3801.10). Those links are written down once, from how the product is made, in config/chains.yaml.

Ladders. Each chain has its own ladder of stages: the traded forms a material takes between the ground and the shop, in the order they are made, one rung per step that customs records can see. Ladders differ in length because the processes do: clothing has four rungs (raw fibre, yarn, fabric, clothing), steel six, chips eight (quartz, silicon metal, polysilicon, wafers, chips, boards, modules and parts, devices). Several products can share a rung, and a chain need not trade every form: potash goes from mine to fertiliser inside one plant. Stage numbers (S1, S2 …) count rungs from the ground within one chain, so they do not compare across chains.

Chokepoint. The stage worth at least $1bn with the highest exporter concentration (HHI, the sum of squared exporter shares), when that HHI is above 0.25; a chain may have none. An HHI above 0.15 is moderate and above 0.25 high.

Shared stages. A stage that two chains share with the same codes is one product, named after the first chain: wafers in chips and solar, PTA and ethylene glycol in plastics and clothing. No value step crosses a shared code.

Feedstocks and use shares. A customs record says what crossed the border, not what it was for. About one barrel of crude in eight becomes petrochemical feedstock (IEA, The Future of Petrochemicals, 2018: about 12% of oil demand), and about two-thirds of nickel goes to stainless steel, not to cathodes. So a chain starts at its first chain-specific product; crude, gas and LPG join as side inputs; and every shared stage carries a use share from a named, dated source outside trade. The share scales the trail and the scenarios but never the map: the map always shows the whole flow, because that is what crossed the border. Where a shared stage has no source, the share reads “unknown” and a scenario stops there. A stage that feeds only its own chain carries no use share: there a scenario passes the whole shortfall down, an explicit upper bound labelled as such (100% pass-through). A scenario’s exposure upstream is the share of an economy’s imports from the closed supplier, not of all it uses, so it overstates the risk where most of the input is made at home.

World trade of a stage is the sum of its flows under the mirror rule. Unit values are shown only where net weight covers at least 80% of value. The ten chains have a page each.

Economies become nodes and flows become directed, weighted edges, exporter → importer, weighted by value. The all-goods network is dense: in 2024, 31,789 flows linked 235 economies, 58% of all possible pairs. A picture of every edge is a hairball, so the views draw a backbone: each economy’s top destinations and top suppliers (k), above a share of world trade (min). Measures are always computed on the full graph.

An economy can import and export the same stage. The Grubel–Lloyd index measures how balanced the two are: GL = 1 − |X − M| / (X + M), for the economy’s exports X and imports M of the stage; 1 means equal, 0 one-way. A chain page marks an economy ↔ two-way when its GL is at least 0.3 and its exports and its imports are each at least 2% of the stage’s world trade (one constant in the builder, TWO_WAY).

Under a stage whose five largest exporters or importers include a two-way trader, one sentence gives the largest one’s imports and exports, the unit value of each side where net weight covers at least 80% of the value, and its three largest suppliers and buyers. The sentence states numbers only; a cause is named only in a hand-written note with a source.

Two-way trade is rarely an error. Four ordinary reasons: one six-digit code can hold goods of very different value (2506.10 holds quartzite lumps and high-purity quartz); goods are imported, refined or assembled and exported again; firms or regions in one economy buy and sell for their own reasons; and under the mirror rule one side can come from the economy’s own reports and the other from its partners’, where origin and consignment, re-exports and misclassification can make one side larger.

Reports disagree. A flow has up to two reports, the importer’s and the exporter’s. The site draws one of them (the mirror rule) and keeps their ratio. Their gap is |importer’s − exporter’s| / the larger; a flow of at least 1% of the product’s world trade whose gap is above 0.5 carries “⚠ reports disagree” in tables and tooltips, never on the map. Imports are valued with freight and insurance and exports without, so gaps of a few per cent are normal.

Split nodes. On chain pages each economy is two half-discs on one scale, imports on the left in teal and exports on the right in gold, each with the area of its value, so the two halves together have the area of total trade. Big exporters, big importers and two-way traders read apart without another legend. The explorer offers the same split as a size.

Maps are coloured by predefined groupings rather than by communities computed from the data, which change from year to year (communities remain an advanced option in the explorer). Flows inside one group take its colour; flows between groups are grey. Each group keeps one colour on every page and in every year.

Regions, the default, are the UN M49 subregions of the Comtrade economy list in eight regions: North America; Latin America and the Caribbean; Europe; the Middle East and North Africa (Northern Africa and Western Asia); Sub-Saharan Africa; Central and South Asia; East and Southeast Asia; Oceania.

Trade blocs and clubs have dated memberships: an economy belongs from the day its membership took effect to its last day as a member. A map shows membership on the last day of the period on screen (31 December for a year, the month’s last day for a month), or today’s when the reader asks. Several blocs can be coloured at once; an economy in two of them takes the first in the reader’s order. Clubs overlap, so they are highlighted one at a time. Readers can also colour their own list of economies.

The table lists every bloc and club with its members on 2 Nov 2026 and the sources of the dates.

Every series passes a ladder of checks before it is published, and the results are published with it: economists want to see the check, not a badge. No commercial data are used: every series and every check on this site comes from a public source. A check that could not run says “not run”, never “pass”.

The last run, on 31 Oct 2026, covered 116 products: 471 pass, 207 watch, 5 fail, 194 info, 267 not run. Every check and its result.

UN Comtrade’s terms of use draw a line at 100,000 records: a free visualisation holding fewer than 100,000 original records needs no licence. Transformed data (new indicators, indices, aggregations) are not copyrighted, while rounding or reformatting is not a transformation. The line counts what a product stores, not what one query returns.

The bundle stores derived records (shares, ranks, ratios) and 0 original records (budget 90,000; UN Comtrade’s line is 100,000). A flow is stored as its share of the product’s world trade, its ranks among the exporter’s destinations and the importer’s suppliers, and its mirror ratio; each product and period carries one world total. Values on screen are share × total. HHI, PageRank, group shares and the strip series are indicators.

There is no bulk path: downloads carry only what a chart shows, there is no API, and the data folder is not listed. Readers who want the records should go to UN Comtrade (comtradeplus.un.org).

Korea’s months that UN Comtrade lacks come from the Korea Customs Service through the public data portal (data.go.kr), which allows reuse with credit to the source. The site credits the Korea Customs Service wherever those months are used, and stores only derived values from them, as for Comtrade.

Conclusion, read against these terms (decision D6, 24 October 2026). The bundle stores derived records only, shares, ranks and ratios, with 0 original records against a budget of 90,000, and offers no bulk download: no API, no listing of the data folder, downloads limited to what a chart shows. On that reading the site sits under UN Comtrade’s line and needs no licence; the records themselves are at comtradeplus.un.org.

Fonts are under the SIL Open Font License 1.1, icons are Lucide (ISC) and the map layout starts from a Natural Earth projection (public domain).

The explorer keeps its state in the address after #, so a view can be shared exactly and an article can link into it. Keys with default values are left out. For example:

/en/explore#t=2024&base=2019&p=C:batteries:cells&c=change&focus=CHN

Chain pages take ?focus=ISO3 to show one economy’s place at every stage (the explorer reads it too); no economy is shown by default.

Cite the data with the build date of the files you used:

United Nations (2026), UN Comtrade Database, Final data, comtradeplus.un.org; derived by trade-flow, trade-flow.datasteg.com, files built 2 Nov 2026.

To cite a view, give its full address, the part after # included, and the date you saw it. The sources quoted on this page are Feenstra et al. (2005) for the mirror rule, Mitikj and Kaushik (2024) for mirror matching, Pierce and Schott (2012) for HS revisions and the IEA (2018) for the use share of crude.