TradeWeave is a free, open-access global trade analytics platform covering 238 countries, 5,022 products, and 30 years of bilateral trade data from BACI/CEPII. It provides interactive visualizations, country profiles, tariff simulations, economic complexity rankings, and trade forecasting tools.

What data sources does TradeWeave use?

TradeWeave combines data from BACI/CEPII (bilateral trade flows), WITS/TRAINS (tariff rates), World Bank WDI (macroeconomic indicators), ESCAP-World Bank (bilateral trade costs), FRED and BLS (freight indices, exchange rates), and FAOSTAT (agricultural trade).

How can I analyze a country's trade profile?

Visit the Country Profiles page and select any of 238 countries to see their export composition, top trade partners, revealed comparative advantage (RCA), Economic Complexity Index (ECI), trade diversification, and 30-year trends.

What is the Economic Complexity Index (ECI)?

The Economic Complexity Index (ECI), developed by Hidalgo and Hausmann, measures the productive knowledge embedded in a country's export basket. Countries that export a diverse set of sophisticated products score higher. TradeWeave ranks all 238 countries by ECI from 1995 to 2024.

Can I simulate the impact of tariff changes?

Yes. TradeWeave's Tariff Simulator lets you model how tariff changes affect import volumes using real product-level demand elasticities and MFN tariff rates for over 1,200 HS4 products.

How far back does the trade data go?

The primary BACI dataset covers 1995-2024 with HS6 product detail. TradeWeave also includes historical trade data spanning over 200 years (1800-present) from academic compilations.

TradeWeave

CtrlK

TradeWeave / Instruments

Loading chart...

Related analyses

Research

Trade questions answered with trade methods

Replications

Canonical papers reproduced and scored

Product Space

The attraction half of the co-occurrence structure

Development Ladder

Entry probability without a memory term

Methodology

Site-wide data sources and conventions

Instrument 03

Counterpoint

This measures when trade happens, not when food grows. Every axis on this page is a shipment month, the month customs cleared the cargo. Fit a phase to all 265,500 monthly export series in the clean segment and the first result is a null: 94.2% of them have no detectable annual cycle at all. Inside the 15,420 that do, the phase carries something the Herfindahl index cannot: it puts the Netherlands and Hong Kong SAR on Chile’s grape calendar rather than their own latitude’s.

Instrument: Circular mean resultant length, Rayleigh test, Mardia circular-linear correlation, on phases from a K = 2 harmonic regression
Borrowed from: Directional statistics, with the unison-versus-counterpoint distinction borrowed from music

TradeWeave measures supply concentration in the cross-section four different ways and none of them has a clock. Two products with the same Herfindahl index over exporters score identically whether their suppliers all ship in March or spread evenly round the year. The missing statistic is the phase: give every exporter a direction on a twelve-month circle and the length of the value-weighted resultant, R̄, says how much the world’s suppliers agree about when.

Three things have to be said before any R̄ on this page is read. The customs record is a clearance date, not a harvest date, so this is a shipment-timing statistic and it is labelled as one throughout. R̄ under the null is not zero: with twenty exporters at uniformly random phases the median R̄ is 0.187 and the 95th percentile is 0.384, so a product at 0.30 is indistinguishable from noise. And the world is mostly aphasic: the median seasonal strength across all 265,500 fitted series is 0.0398. The atlas is the 15,420 series that clear a 0.25 gate, not the whole record. The sonification the original proposal asked for is cut: by its own test it is decoration unless a forced-choice listening test passes, and an unpassed self-test does not ship.

Kill condition

The load-bearing claim here is the entrepot diagnostic: that an exporter shipping on the opposite hemisphere’s clock is detectable from timing alone. The run that decides it is the seasonality-gate sweep at the canonical latitude clock (Figure 13). The measured false-positive floor is the share of flagged exporter-product lines falling outside HS chapters 01 to 24, where no harvest clock can physically exist. If tightening the gate fails to push that share down, or fails to concentrate the flag on the entrepots, then the flag is noise wearing a phase and this page is retired with a dated banner and a ledger entry.

The run: at gate 0.00 the diagnostic flags 2,284 of 10,844 lines and 72.7% of the flags are outside chapters 01 to 24, which is mostly noise. At 0.25 it flags 99 of 1,676 with 33.3% outside. At 0.50 it flags 30 of 684 with 6.7% outside. Over the same tightening the Netherlands’ share of flagged value rises from 0.070 to 0.266 and Hong Kong SAR’s from 0.198 to 0.491, while Russia’s collapses from 0.270 to 0.000 and China’s from 0.126 to 0.000. The diagnostic survives. It survives with a false-positive floor of 33.3% at the canonical setting, which is printed everywhere the flag is used.

This sweep has already killed one version of this page. The first pass, with one criterion and no seasonality gate, ranked Russia top of the entrepot table and returned Chinese sports footwear as the largest flagged line in the world. Neither is transshipment. They are a noise phase and a Christmas-retail clock mistaken for a harvest clock.

Series fitted265,500MEASURED reporter × HS6, ≥ 60 months

With a detectable phase5.81%MODELED 15,420 series

Median product R̄0.282–0.597MODELED unweighted to value weighted

HHI vs R̄r = 0.351MODELED 12.3% of the variance

Netherlands, grapes5.03 moMODELED from its own hemisphere’s clock

The instrument

For each reporter and HS6 code with at least 60 distinct months in the clean segment, regress log monthly export value on a linear trend and K = 2 harmonic pairs at the annual and semi-annual frequencies. The first-harmonic amplitude R₁ is the size of the annual swing in log points and its phase is the peak shipment month. Seasonal strength is the share of the detrended variance the harmonics explain; a series is called seasonal when that clears 0.25. At product level, take the top 20 exporters by value, place each at its fitted peak month as a unit phasor, weight by export value, and R̄ is the length of the resultant. R̄ near 1 is unison, every supplier shipping at the same point of the year. R̄ near 0 is counterpoint, the world covered round the clock.

A phase on the (0, 12] circle is written as a month number: 1 is the January mark, 12 the December mark, and 3.23 sits between Mar and Apr. A value below 1, like Peru’s 0.24, lands just before the January mark, which is late December. Two specifications are defensible for the phase, the detrended harmonic fit and the value-weighted circular mean of raw monthly value, and both are shipped on every table here because at product level they disagree far more than is comfortable (Figure 6).

Figure 1

The segment, and everything dropped before a single fit

Metric	Value	What it is
rows_clean_export_segment	45,874,182	flowCode='X', period 200001..202308
reporter_codes	167	distinct reporterCode, includes aggregates with NULL ISO
reporter_iso3	163	distinct reporterISO after dropping NULL
hs6_codes	6,688	distinct 6-digit cmdCode
rows_null_reporter_iso	1,271,164	EU-28, ASEAN and Faeroe Isds carry NULL reporterISO
rows_nonpositive_value	332	dropped before the log fit
total_value_usd	269.56tn USD	primaryValue is USD not thousands; rounded to the nearest million
duplicate_reporter_period_flow_cmd_keys	0	SOURCE.md claims zero for this segment; re-verified here
series_ge60_months_all_reporter_codes	275,460	spec figure, keyed on reporterCode without the NULL-ISO drop
series_ge60_months_drop_null_iso	265,503	method step 1 applied as written (drop NULL reporterISO)
series_ge60_months_drop_null_iso_positive	265,500	the panel actually fitted

The fitted panel is 265,500 series, not the 275,460 a reporter-code count would give. MEASURED The gap is the NULL-ISO drop: 1,271,164 rows carry no reporter ISO3 and they come from three reporter codes, not from aggregates alone. Two are aggregates (EU-28 and ASEAN) and one, the Faeroe Islands, is a real territory that costs 290 rows to drop. Sudan appears under two reporter codes, 729 and 736 (Fmr Sudan), which is why every series here is keyed on reporterCode and not on the ISO3.

The first result is a null

The pitch for this page assumed world trade has a clock and that the job was to read it. It mostly does not. Of 265,500 fitted series, 5.81% clear the 0.25 seasonal-strength gate, covering 2,734 HS6 codes and 125 reporters. The median series puts 0.0398 of its detrended variance into the annual and semi-annual harmonics. Any framing in which most of world trade carries a detectable annual phase is false, and the atlas below is an atlas of 15,420 series.

Figure 2

How much of world trade has a clock at all

Figure 3

R̄ under the null is not zero

Exporters	Null R̄ p50	p90	p95	p99	Analytic R̄ at p = 0.05
5	0.385	0.672	0.754	0.885	0.774
8	0.301	0.535	0.604	0.725	0.612
10	0.267	0.477	0.540	0.656	0.547
12	0.243	0.438	0.496	0.596	0.500
15	0.217	0.387	0.439	0.543	0.447
20	0.187	0.336	0.384	0.474	0.387
25	0.168	0.302	0.343	0.423	0.346
30	0.153	0.276	0.314	0.388	0.316

With 20 exporters at uniformly random phases, the median R̄ is 0.187 and the 95th percentile is 0.384. MODELED This is why 3,578 products of 5,613 score above 0.5 but only 2,136 reach a Monte-Carlo Rayleigh p below 0.05, and only 1,531 reach it on the conservative analytic test. Read R̄ against the null column for the right exporter count, never against zero.

The dial

Grapes, HS 080610, are the canonical case and they do not behave the way the pitch expected. As a product they are counterpoint, not unison: R̄ = 0.1459 with a Monte-Carlo Rayleigh p of 0.8445 and a bootstrap 95% interval of [0.073, 0.696] that straddles the null 95th percentile of 0.539. The world grape clock is not detectable as a single harvest, which is the point: 132 exporters cover the calendar between them. What the dial shows is not concentration. It is who ships when, and which spokes sit where their latitude says they should not.

Figure 4

Grapes, HS 080610: fourteen exporters on a twelve-month dial

Figure 5

The same grapes with no fit in them at all

Figure 6

How far apart the two defensible phase specifications land

Comparison, harmonic R̄ against circular-mean R̄	Value
Products compared	5,613
Pearson correlation	0.618
Spearman correlation	0.648
Median absolute gap in R̄	0.107
90th percentile gap	0.337
Share differing by more than 0.10	52.1%
Share differing by more than 0.30	13.3%
Share crossing R̄ = 0.5 between specifications	25.6%
Median gap where R̄ below 0.3	0.196
Median gap where R̄ above 0.7	0.061
Log against asinh, Pearson	0.924
Log against asinh, median absolute gap	0.030

At product level the harmonic R̄ and the circular-mean R̄ correlate at Pearson 0.618 only, and 25.6% of products cross the R̄ = 0.5 line depending on which one you pick. MODELED The disagreement is concentrated exactly where it should be: median gap 0.196 among weakly seasonal products (R̄ below 0.3) against 0.061 among strongly seasonal ones (R̄ above 0.7). On genuinely seasonal series the two specs are interchangeable. On weak ones they are not, because the circular mean of raw monthly value picks up month length and common-calendar effects that the detrended fit removes. This is the reason both columns ship on every table here.

Computed across the same 5,613 products under both phase specifications. The log-versus-asinh comparison in the same table is far tamer, Pearson 0.924 with a median absolute R̄ gap of 0.030: the transform choice matters much less than the phase-estimator choice.

Specification comparison: data/parquet/instruments_counterpoint_spec_compare.parquet, built from the Rbar, Rbar_circ and Rbar_asinh columns of instruments_counterpoint_products.parquet.

Figure 7

The counterpoint finder: who covers you when you are out of season

Exporter	Its peak month	Closest counterpoint	Counterpoint peak	Separation (months)
Chile	3.14	Italy	9.23	5.91
Italy	9.23	Chile	3.14	5.91
United States	9.33	Australia	3.44	5.89
Netherlands	2.64	Italy	9.23	5.41
South Africa	1.70	Egypt	7.11	5.41
China	9.49	Australia	3.44	5.95
Peru	0.14	Egypt	7.11	5.03
Spain	9.28	Chile	3.14	5.87
Hong Kong SAR	3.01	Italy	9.23	5.78
Australia	3.44	Turkey	9.48	5.96
Egypt	7.11	South Africa	1.70	5.41
India	2.36	Italy	9.23	5.13
Turkey	9.48	Australia	3.44	5.96
Brazil	9.60	Australia	3.44	5.84

Figure 8

Unison and counterpoint at the top of world trade

HS6	Product	R̄ harmonic	R̄ circular	Peak month	MC Rayleigh p	Analytic p	Fitted value
210500	Ice cream and other edible ice: whether or not containing coc…	0.9932	0.9937	5.39	0.000	0.000	$50.3bn
220410	Wine: sparkling	0.9899	0.9775	9.64	0.000	0.063	$107.9bn
430310	Furskin articles: apparel and clothing accessories	0.9818	0.9598	8.93	0.000	0.075	$42.1bn
610433	Jackets: women's or girls', of synthetic fibres, knitted or c…	0.9740	0.9487	8.74	0.000	0.219	$22.6bn
120100	Soya beans: whether or not broken	0.9720	0.9694	6.31	0.166	0.392	$96.3bn
180690	Chocolate and other food preparations containing cocoa: n.e.s…	0.9710	0.9645	10.36	0.000	0.000	$171.6bn
030541	Fish: Pacific salmon (oncorhynchus spp.), Atlantic salmon (sa…	0.9690	0.9073	11.39	0.000	0.006

The blind spot the page was built for

If R̄ were recoverable from the cross-section, this instrument would be redundant. It is not. Across 5,613 products, exporter HHI and R̄ correlate at Pearson 0.351, Spearman 0.309, so the concentration measure this site already computes explains 12.3% of the variance in the timing measure it does not. Inside a single narrow HHI band the entire R̄ range is present.

Figure 9

Identical cross-sectional concentration, opposite calendars

HHI band	Products	Median HHI	Min R̄	p10	Median R̄	p90	Max R̄	R̄ IQR
(0.0, 0.05]	21	0.0469	0.1663	0.229	0.6040	0.839	0.9135	0.402
(0.05, 0.1]	1,201	0.0822	0.0286	0.220	0.5231	0.805	0.9932	0.316
(0.1, 0.15]	1,600	0.1227	0.0233	0.216	0.5368	0.830	0.9625	0.341
(0.15, 0.2]	1,021	0.1718	0.0083	0.265	0.5859	0.859	0.9703	0.330
(0.2, 0.3]	931	0.2365	0.0245	0.320	0.6383	0.886	0.9925	0.317
(0.3, 0.5]	606	0.3729	0.0881	0.451	0.7550	0.928	0.9974	0.268
(0.5, 1.0]

Figure 10

The twin: same HHI to three decimals, R̄ 0.96 against 0.12

HS 220210, Waters: including mineral and aerated, containing added sugar or other sweetening matter or flavoured. HHI 0.0658, R̄ 0.964 harmonic / 0.961 circular, $95.3bn.

The byproduct: a clock in the wrong hemisphere

Grapes have a hemisphere structure the fit recovers without being told about it. Among the 43 northern and 5 southern grape exporters outside the tropics, the value-weighted mean peak months are 9.374 and 2.884, 5.51 months apart, with a Mardia circular-linear correlation between port latitude and phase of 0.328. That is the reference clock. An exporter whose own phase sits closer to the opposite hemisphere’s clock than to its own, by a margin, is flagged. The Netherlands ships grapes at month 2.645 from a vessel-weighted port latitude of 52.0 degrees north, 5.03 months from its own hemisphere’s clock. Hong Kong SAR ships at 3.009 from 22.3 degrees, 5.51 months out, which is the maximum the 5.51-month hemisphere separation allows.

Figure 11

Grapes: port latitude against fitted shipment phase

Figure 12

What the flag actually catches, including what it gets wrong

Exporter	Port lat	HS6	Product	Peak month	Strength	Margin (mo)	Value	Harvest possible
Netherlands	52.0	080610	Fruit, edible: grapes, fresh	2.64	0.95	5.03	$9.72bn	HS 01-24
China	31.2	730640	Steel, stainless: tubes and pipes, welded, of circular cross-…	8.24	0.44	2.05	$4.89bn	no, false positive
China	31.2	230400	Oil-cake and other solid residues: whether or not ground or i…	5.95	0.27	2.68	$4.44bn	HS 01-24
Hong Kong SAR	22.3	080610	Fruit, edible: grapes, fresh	3.01	0.67	5.51	$4.24bn	HS 01-24
Hong Kong SAR	22.3	640411	Sports footwear: tennis shoes, basketball shoes, gym shoes, t…	11.26	0.25	4.49	$3.51bn	no, false positive
Netherlands	52.0	080810	Fruit, edible: apples, fresh	6.36	0.69	4.79	$3.44bn	HS 01-24
Netherlands	52.0	080510

Exploratory reading

The obvious reading of the Netherlands and Hong Kong SAR on the southern clock is transshipment: fruit grown in Chile, Peru or South Africa, cleared through Rotterdam or Hong Kong, and re-exported with the southern harvest calendar still attached to it. The data here cannot identify that. It identifies a clearance calendar in the wrong hemisphere, which is consistent with re-export, with cold storage, with contract timing and with a domestic season that happens to be counter-cyclical. What makes the re-export reading attractive rather than proven is that it survives the tightening that destroys the alternatives: at the strict seasonality gate the flag concentrates on exactly the two countries known for entrepot trade and vanishes for Russia and China. That is suggestive. It is not a customs finding and this page does not treat it as one.

Every free choice, swept

Seven things on this page could have been chosen differently. Each one gets a table. The largest sensitivity is not the harmonic order or the estimation window: it is the weighting and the exporter cut, which move the median product R̄ from 0.282 to 0.597, a factor of 2.1.

Figure 13

Sweep 1, the entrepot clock: four parameters, and the kill-condition run

Seasonality gate	Clock products	Lines	Flagged	Flagged share	Outside HS 01-24	NLD	HKG	RUS	CHN
0.00	299	10,844	2,284	21.1%	72.7%	0.070	0.198	0.270	0.126
0.05	299	5,699	800	14.0%	60.5%	0.120	0.177	0.099	0.113
0.10	289	3,481	344	9.9%	51.5%	0.153	0.203	0.079	0.063
0.25	193	1,676	99	5.9%	33.3%	0.190	0.357	0.054	0.052
0.50	113	684	30	4.4%	6.7%	0.266	0.491	0.000	0.000

Separation ≥	Hemisphere R̄ ≥	Clock products	Lines

Figure 14

Sweep 2, exporter count and weighting: the largest free parameter on the page

Top N exporters	Weighting	Median R̄	Mean R̄	Share R̄ > 0.5	Share MC p < 0.05	Grapes R̄
5	value weighted	0.6992	0.6642	75.5%	15.4%	0.2123
10	value weighted	0.6313	0.6095	68.0%	23.4%	0.1896
15	value weighted	0.6091	0.5904	65.2%	26.1%	0.1518
20	value weighted	0.5967	0.5817	63.7%	27.3%	0.1459
30	value weighted	0.5880	0.5745	62.7%	28.3%	0.1454
50	value weighted	0.5850	0.5713	62.3%	28.6%	0.1453
all	value weighted	0.5845	0.5709	62.3%	28.6%	0.1454
5	unweighted	0.5762	0.5733	61.3%	27.8%	0.2584
10

Figure 15

Sweep 3, number of harmonic pairs K

K	Series	Median R₁	Median strength	Share above gate	Median peak shift vs K = 2	p90 shift	Median product R̄	Grapes R̄
1	265,500	0.1862	0.0192	3.94%	0.024	0.143	0.5983	0.1459
2	265,500	0.1864	0.0398	5.81%	0.000	0.000	0.5967	0.1459
3	265,499	0.1864	0.0579	7.45%	0.011	0.112	0.5980	0.1451
4	265,487	0.1866	0.0731	8.84%	0.018	0.168	0.5968	0.1452

Phase is essentially invariant to K: the median peak month moves by under 0.024 months across all four settings. MODELED What K does move is the seasonal-strength gate: median strength runs from 0.0192 at K = 1 to 0.0731 at K = 4, mechanically, because more harmonics explain more variance. Product R̄ is unaffected to three decimals.

A fit is kept only if the scaled normal-equation matrix has a minimum eigenvalue above 1e-9 and the series has at least 2K + 1 distinct months of the year. All 265,500 series pass at K = 2; 13 fail at K = 3 and 26 at K = 4, which is why the series count is not constant down the column.

Figure 16

Sweep 4, minimum distinct months required to fit a series

Min months	Series	Reporters	HS6 fitted	Median strength	Median R₁	Products	Median product R̄	Grapes R̄
36	314,135	146	6,071	0.0455	0.2131	5,834	0.5936	0.1458
48	286,580	139	6,039	0.0420	0.1984	5,734	0.5957	0.1458
60	265,500	133	6,009	0.0398	0.1864	5,613	0.5967	0.1459
72	248,460	128	5,942	0.0384	0.1768	5,532	0.5993	0.1459
96	212,628	115	5,381	0.0361	0.1618	5,170	0.6038	0.1459
120	174,695	104	5,325	0.0340	0.1498	5,054	0.6093	0.1673

The threshold trades panel size against fit quality and buys almost nothing either way. Grapes R̄ is stable to four decimals from 36 to 96 months and only moves at 120, where the panel has lost a third of its series.

Figure 17

Sweep 5, the seasonal-strength gate

Gate	Series passing	Share of panel	Median R₂/R₁	p90 R₂/R₁	Share with R₂ > R₁	Median R₁ among passers
0.05	109,148	41.11%	0.7066	2.0964	33.13%	0.2941
0.10	50,743	19.11%	0.5824	1.7012	24.76%	0.3584
0.15	30,068	11.33%	0.5017	1.3949	18.46%	0.4222
0.20	20,827	7.84%	0.4528	1.2058	14.53%	0.4887
0.25	15,420	5.81%	0.4201	1.1005	12.04%	0.5558
0.30	11,891	4.48%	0.3948	0.9993	9.97%	0.6181
0.40	7,303	2.75%	0.3578	0.8410	6.98%	0.7498
0.50	4,540	1.71%	0.3291	0.7398	4.91%

Figure 18

Sweep 6, estimation window

Window	Periods	Series	Median strength	Products	Median product R̄	Median peak shift vs full	Grapes R̄
full	200001 to 202308	265,500	0.0398	5,613	0.5967	0.000	0.1459
w2010	201001 to 202308	255,190	0.0441	5,312	0.5960	0.000	0.1479
w2015	201501 to 202308	216,668	0.0600	5,176	0.5796	0.578	0.1692

Restricting the fit to 2010 onward changes nothing; restricting it to 2015 onward moves the median peak month by 0.578 months. MODELED The 2015 window also raises median seasonal strength from 0.0398 to 0.0600, which is what a shorter window does to any harmonic fit: less time for the seasonal pattern to drift. The canonical choice is the full segment because it is the longest window with zero duplicate keys.

All three windows end at 2023m8, the last month of the clean segment. The full window starts in 2000m1, when only 21 reporters were filing monthly data, which is why the fixed-panel correction in Figure 20 exists.

Window sweep: data/parquet/instruments_counterpoint_sweep_window.parquet.

Figure 19

Sweep 7, the months the log specification deletes

asinh scale θ	Median log R₁	Median asinh R₁	Amplitude ratio, all series	Amplitude ratio, affected series	Median peak shift (mo)	Share shifting > 1 mo
series median positive month	0.1864	0.0971	0.596	0.504	0.369	23.6%
series median positive month / 10.0	0.1864	0.1635	0.978	0.878	0.185	21.0%
series median positive month / 100.0	0.1864	0.2000	1.000	1.127	0.155	25.2%
1.0 USD	0.1864	0.2578	1.000	1.677	0.213	32.7%

Share of months the log spec deletes	Series	Amplitude ratio	Median peak shift (mo)	Share shifting > 1 mo
(-0.001, 0.0]	116,420	0.656	0.190	9.4%
(0.0, 0.02]	27,966	0.603	0.359	18.9%
(0.02, 0.05]	14,745	0.552	0.503	27.3%
(0.05, 0.1]	17,251	0.553	0.537

Figure 20

Who was filing, and what a moving reporter set does to a pooled statistic

What did not survive the computation

The premise that world trade has a clock. Median seasonal strength is 0.0398 and 5.81% of series clear the gate. The atlas is 15,420 series of 265,500, and any framing implying most trade carries a detectable annual phase is false.
Grapes as a concentrated harvest. At product level grapes are R̄ = 0.1459 with Monte-Carlo p = 0.8445. They are the counterpoint case, not evidence of synchronisation. The world covers grapes round the calendar; that is the finding, and it is the opposite of what the pitch assumed.
The first HHI twin. HS 051199, animal products: n.e.s. in chapter 5, scored R̄ 0.029 on the harmonic specification and 0.670 on the circular-mean one. It was replaced with a pair that survives both, at a worst-case gap of 0.8422.
The one-criterion entrepot flag. Without a seasonality gate it put Russia at the top of the entrepot ranking and returned Chinese sports footwear as the largest flagged line. Four gated parameters and a printed false-positive floor of 33.3% replaced it.
The series count. 275,460 is the count before the NULL-ISO drop that the method itself specifies. The fitted panel is 265,500.
The sonification. Cut outright. By the proposal’s own test it is decoration unless a forced-choice listening test passes, and an unpassed self-test is not something this section ships.

The strongest objection to this page

Customs records the month of clearance, not the month of harvest. Contracts, cold storage and demand spikes around Chinese New Year or Ramadan all move the recorded peak away from any physical cycle, so this atlas is not measuring agriculture.

Secondly, reporter coverage in the monthly file is unstable, so any pooled world statistic drifts with whoever happened to be filing that year.

The first objection is conceded rather than argued with, because the concession costs nothing. This page measures when trade happens, not when food grows: the dek says so, every axis label says shipment month, and the word harvest appears only where the hemisphere reference clock is being defined. Both results that carry the page are about shipment timing by construction. R̄ is a supply-security statistic about when cargo actually arrives, which is the operationally relevant quantity for anyone holding a contract, and the entrepot diagnostic works precisely because clearance timing and growing season come apart: it is that gap the flag is reading.

The objection does bite in one place and the page prints the damage. Of the 99 flagged lines at the canonical setting, 33.3% fall outside HS chapters 01 to 24, where no harvest clock can physically exist. Those are demand calendars, Christmas retail and contract cycles, exactly the mechanism the objection names, and at a strict gate the share falls to 6.7% (Figure 13). The flag is never claimed to prove transshipment.

The second objection is answered mechanically. Every pooled statistic here is computed on a fixed panel of 61 reporters that filed at least 11 months in every year from 2010 to 2022, and the panel’s membership, the unbalanced comparison and the size of the distortion are all published in Figure 20: the unbalanced panel overstates within-series monthly concentration in all 13 years, by 1.7% to 34.4%. No series is fitted outside the zero-duplicate segment ending 2023m8, and the duplicate count for that segment was re-verified here rather than taken on trust: 0 keys (Figure 1).

What the objection does not touch is the twin. HS 220210 and HS 200190 have the same exporter Herfindahl index to three decimals and R̄ of 0.964 against 0.118. Whatever the phase is measuring, clearance calendars or growing seasons or contract cycles, it is not recoverable from the concentration statistics this site already computes, and that was the claim.

Every figure on this page reads a precomputed table under data/parquet/instruments_counterpoint_*.parquet; the build script that produced them is data/build_instruments_counterpoint.py. The full atlas (265,500 rows) is not read by this page, which uses the gated subset of 15,420 series wherever it needs series-level rows. primaryValue is USD not thousands; rounded to the nearest million.

Exporter	Peak month, harmonic	Peak month, circular mean	Seasonal strength	R₁ (log pts)	Value in window
Chile	3.143	3.059	0.842	4.509	$24.71bn
Italy	9.231	9.338	0.922	2.737	$16.10bn
United States	9.328	9.458	0.933	1.212	$11.91bn
Netherlands	2.645	2.701	0.947	1.202	$9.72bn
South Africa	1.700	1.597	0.937	3.366	$7.25bn
China	9.487	9.373	0.688	2.431	$6.22bn
Peru	0.145	0.179	0.853	3.856	$6.10bn
Spain	9.275	9.023	0.865	1.486	$6.04bn
Hong Kong SAR	3.009	3.633	0.668	1.671	$4.24bn
Australia	3.438	3.499	0.887	4.075	$3.58bn
Egypt	7.111	5.910	0.758	4.501	$3.14bn
India	2.360	2.639	0.831	4.409	$3.10bn
Turkey	9.482	9.235	0.892	4.580	$3.02bn
Brazil	9.603	10.126	0.654	1.742	$2.26bn

Italy’s grape peak falls 5.97 to 6.09 months after Chile’s, which is as close to antiphase as a twelve-month circle allows. MODELED Spoke direction is the fitted peak shipment month, length is seasonal strength, thickness is share of world grape value. The accent arrow is the product resultant, R̄ = 0.1459 pointing at month 0.35, and it does not even reach the dashed ring, which is the Monte-Carlo null 95th percentile at 0.539. Five spokes are drawn in accent: the Netherlands, Hong Kong SAR, Germany, Britain and Denmark, all northern ports shipping on the southern clock. The blue rim ticks are the same exporters under the other phase specification.

The 14 largest grape exporters by fitted-window value among the 15,420 series that clear the 0.25 gate; 46 grape exporters clear it in total and 132 appear in the segment. The resultant is the canonical product statistic over the top 20 exporters, so it is not the resultant of the 14 spokes drawn. Rim ticks are the value-weighted circular mean of raw monthly value, the second defensible specification; the two agree to a median of 0.070 months across the 12 largest exporters and disagree worst on Egypt at 1.21 months. Phase convention: the tick labelled Jan is exactly month 1.0, so a phase of 0.24 falls just before it, in late December. A spoke label is dropped where it would print on top of another; the table carries all fourteen exporters either way.

UN Comtrade raw monthly bulk, clean export segment 2000m1 to 2023m8, data/parquet/comtrade_monthly.parquet (flowCode = X, primaryValue in USD, not thousands; Germany 2022m12 is absent from the source bulk, one corrupt gzip, documented in comtrade_monthly.SOURCE.md). Series fits: data/parquet/instruments_counterpoint_atlas_seasonal.parquet (peak_month, seasonal_strength, circ_peak_month). Product resultant: data/parquet/instruments_counterpoint_products.parquet. HS6 names: data/parquet/products_all.parquet, matched through HS22, HS17, HS12, HS07, HS02, HS96, HS92 in that order with comtrade_monthly.cmdDesc as the last fallback; every product row carries label_revision and label_source.

Query

SELECT iso3, peak_month, seasonal_strength, R1, circ_peak_month, circ_Rbar,
       value_usd, n_obs, first_period, last_period
FROM read_parquet('data/parquet/instruments_counterpoint_atlas_seasonal.parquet')
WHERE cmdCode = '080610'
ORDER BY value_usd DESC
LIMIT 14

Among products above $20bn of fitted export value, R̄ runs from 0.9932 for ice cream down to 0.0245 for footwear soles and heels. MODELED The unison list is a list of calendars, not of harvests: ice cream in May, sparkling wine and fur apparel and synthetic jackets and wristwatches in the northern autumn shipping season, chocolate preparations in October. The counterpoint list is industrial: toluene, polyester staple fibre, plastic furniture fittings, ignition parts. Both p-value columns are shown because the analytic test is the conservative one.

Fitted export value is the top-20-exporter total in the clean segment, not world trade in the product; the world column beside it is the full exporter set. Peak month is the value-weighted circular mean of the exporters’ fitted phases. Where the two specifications disagree, both R̄ values are printed side by side. MODELED throughout: the top-20 cut and the value weighting are free choices and their sweep is Figure 14.

UN Comtrade raw monthly bulk, clean export segment 2000m1 to 2023m8, data/parquet/comtrade_monthly.parquet (flowCode = X, primaryValue in USD, not thousands; Germany 2022m12 is absent from the source bulk, one corrupt gzip, documented in comtrade_monthly.SOURCE.md). Product statistics: data/parquet/instruments_counterpoint_products.parquet. HS6 names: data/parquet/products_all.parquet, matched through HS22, HS17, HS12, HS07, HS02, HS96, HS92 in that order with comtrade_monthly.cmdDesc as the last fallback; every product row carries label_revision and label_source.

Query

SELECT cmdCode, product_name, label_revision, n_exporters, n_exporters_all, n_eff,
       value_usd, world_value_usd, hhi_exporters, Rbar, Rbar_circ,
       mean_peak_month, mean_peak_month_circ, rayleigh_p, rayleigh_p_mc,
       null_Rbar_p95, Rbar_boot_lo, Rbar_boot_hi
FROM read_parquet('data/parquet/instruments_counterpoint_products.parquet')
WHERE value_usd > 20e9
ORDER BY Rbar DESC
LIMIT 8

/* and the mirror image */

SELECT cmdCode, product_name, label_revision, n_exporters, n_exporters_all, n_eff,
       value_usd, world_value_usd, hhi_exporters, Rbar, Rbar_circ,
       mean_peak_month, mean_peak_month_circ, rayleigh_p, rayleigh_p_mc,
       null_Rbar_p95, Rbar_boot_lo, Rbar_boot_hi
FROM read_parquet('data/parquet/instruments_counterpoint_products.parquet')
WHERE value_usd > 20e9
ORDER BY Rbar ASC
LIMIT 8

HS 220210 and HS 200190 have exporter HHIs of 0.0658 and 0.0639, a gap of 0.0019, and R̄ of 0.964 against 0.118. MODELED Flavoured waters are shipped by everyone at once: all twelve of the largest exporters peak inside a 1.13-month window centred between Jun and Jul. Vinegar-preserved vegetables have the same supplier concentration and no shared calendar at all, with peaks from Netherlands at 0.92 to Poland at 11.87.

This twin was re-selected. The first pass paired HS 220210 with HS 051199, animal products: n.e.s. in chapter 5, at R̄ 0.029, and that pairing dies under the second phase specification, where 051199 scores 0.670 and flips from counterpoint to near-unison. The pair shown survives both: worst-case R̄ gap across specifications 0.8422 against a headline gap of 0.8456. Spoke length is seasonal strength; the accent arrow is the product resultant and the dashed ring the Monte-Carlo null 95th percentile for that product. A max-minus-min window is only meaningful when the peaks do not wrap the circle, which is why it is quoted for the unison product only.

UN Comtrade raw monthly bulk, clean export segment 2000m1 to 2023m8, data/parquet/comtrade_monthly.parquet (flowCode = X, primaryValue in USD, not thousands; Germany 2022m12 is absent from the source bulk, one corrupt gzip, documented in comtrade_monthly.SOURCE.md). Twin selection: data/parquet/instruments_counterpoint_hhi_twins.parquet; exporter phases: data/parquet/instruments_counterpoint_twin_profile.parquet. HS6 names: data/parquet/products_all.parquet, matched through HS22, HS17, HS12, HS07, HS02, HS96, HS92 in that order with comtrade_monthly.cmdDesc as the last fallback; every product row carries label_revision and label_source.

Query

SELECT cmd_unison, name_unison, hhi_unison, Rbar_unison, Rbar_circ_unison, peak_unison,
       value_unison_usd, cmd_counterpoint, name_counterpoint, hhi_counterpoint,
       Rbar_counterpoint, Rbar_circ_counterpoint, peak_counterpoint, value_counterpoint_usd,
       hhi_gap, Rbar_gap, Rbar_gap_worst_case
FROM read_parquet('data/parquet/instruments_counterpoint_hhi_twins.parquet')
WHERE cmd_unison = '220210' AND cmd_counterpoint = '200190'

/* exporter phases behind the two dials */

SELECT DISTINCT cmdCode, role, iso3, peak_month, seasonal_strength
FROM read_parquet('data/parquet/instruments_counterpoint_twin_profile.parquet')
ORDER BY cmdCode, peak_month

The cloud separates cleanly into a southern-latitude group shipping early in the year and a northern group shipping in autumn, and five northern ports sit with the southern group. MODELED Chile, South Africa and Australia are at bottom left: southern latitudes, phases in the first quarter. Italy, Spain, Turkey, the United States and China are at upper right, northern latitudes with autumn phases. The five accent points break the pattern. The Netherlands at 52.0 degrees north sits at top left with Germany, Britain and Denmark, and Hong Kong SAR at 22.3 degrees north sits below them, all of them on the southern clock.

Points are grape exporters whose fit clears the 0.25 seasonal-strength gate and whose vessel-weighted port latitude is at least 20 degrees from the equator, the diagnostic’s own admission rule; India and Peru are excluded by the tropics rule, not by their phases. Marker area is share of world grape value in the 2015m1 to 2023m8 window. Labels are shown for exporters above 1% of world value and for every flagged exporter. Latitude is a port-traffic weighted mean, which is not the same as a country centroid.

Query

SELECT g.iso3, g.port_lat, g.harmonic_peak_month, g.harmonic_strength, g.world_share,
       (e.iso3 IS NOT NULL) AS flagged,
       CASE WHEN g.world_share >= 0.01 OR e.iso3 IS NOT NULL THEN g.iso3 END AS label
FROM read_parquet('data/parquet/instruments_counterpoint_grapes.parquet') g
LEFT JOIN (SELECT iso3 FROM read_parquet('data/parquet/instruments_counterpoint_entrepot_lines.parquet') WHERE cmdCode = '080610') e
  ON e.iso3 = g.iso3
WHERE g.harmonic_strength >= 0.25 AND abs(g.port_lat) >= 20
ORDER BY g.world_share DESC

Country	Port lat	Lines	Flagged	Value in clock products	Flagged share of value
Hong Kong SAR	22.3	30	8	$29.9bn	0.3567
Portugal	39.1	38	4	$6.0bn	0.2126
Korea, Rep.	35.7	15	2	$1.7bn	0.1986
Netherlands	52.0	73	8	$107.1bn	0.1903
Belgium	51.3	61	3	$29.2bn	0.1358
Bulgaria	42.7	44	3	$6.9bn	0.1014
Mexico	20.8	27	4	$35.0bn	0.0753
Japan	35.1	41	2	$10.7bn	0.0663
Canada	47.3	37	6	$12.0bn	0.0554
Russia	51.2	67	6	$3.7bn	0.0540
China	31.2	72	4	$228.2bn	0.0522

At the canonical setting the diagnostic flags 99 exporter-product lines, and 33.3% of them fall outside HS chapters 01 to 24, where no harvest clock can exist. MODELED That is the measured false-positive floor and it is not an estimate: Chinese stainless steel tube, Hong Kong sports footwear and Chinese suture needles are in the flagged list and none of them has a growing season. The largest flagged line in the world is Netherlands grapes at $9.7bn. The country table underneath ranks by the share of a country’s value inside latitude-clock products that gets flagged: Hong Kong SAR heads it at 0.3567, with the Netherlands fourth on that share while carrying much the largest flagged line in absolute value.

Canonical setting: hemisphere separation at least 4.5 months, both hemisphere resultants at least 0.5, exporter seasonal strength at least 0.25, margin at least 1.0 month, latitude reference at least 20 degrees from the equator. Country rows are restricted to countries with at least 5 lines inside qualifying products. The flag is a timing coincidence, not a customs finding: it says an exporter’s clearance calendar matches the other hemisphere, and nothing more. The sweep over all four parameters is Figure 13.

UN Comtrade raw monthly bulk, clean export segment 2000m1 to 2023m8, data/parquet/comtrade_monthly.parquet (flowCode = X, primaryValue in USD, not thousands; Germany 2022m12 is absent from the source bulk, one corrupt gzip, documented in comtrade_monthly.SOURCE.md). Flagged lines: data/parquet/instruments_counterpoint_entrepot_lines.parquet; country scores: data/parquet/instruments_counterpoint_entrepot_countries.parquet. Port latitudes: IMF PortWatch, data/parquet/portwatch_ports.parquet (lat, vessel_count_total), aggregated to a vessel-weighted mean latitude per country. HS6 names: data/parquet/products_all.parquet, matched through HS22, HS17, HS12, HS07, HS02, HS96, HS92 in that order with comtrade_monthly.cmdDesc as the last fallback; every product row carries label_revision and label_source.

Query

SELECT cmdCode, iso3, product_name, port_lat, peak_month, seasonal_strength, value_usd,
       own_hemi_peak, opp_hemi_peak, hemisphere_separation_months, mardia_r_cl,
       margin_months, is_agri
FROM read_parquet('data/parquet/instruments_counterpoint_entrepot_lines.parquet')
ORDER BY value_usd DESC
LIMIT 12

/* country scores */

SELECT iso3, n_lines, n_flagged, value_usd, flagged_value_share, port_lat, vessels
FROM read_parquet('data/parquet/instruments_counterpoint_entrepot_countries.parquet')
WHERE n_lines >= 5
ORDER BY flagged_value_share DESC
LIMIT 11

Tightening the exporter seasonality gate from 0.00 to 0.50 cuts the share of flags that cannot be a harvest from 72.7% to 6.7%, and moves the flagged value toward the two entrepots and away from Russia and China. MODELED The top table sweeps the gate at the canonical clock; the bottom sweeps the clock’s own definition, the hemisphere separation minimum and the minimum hemisphere resultant. Both move in the same direction: strictness buys precision and costs coverage, from 2,284 flagged lines down to 30. The canonical setting is the middle row and its false-positive floor of 33.3% is printed wherever the flag appears.

Top table: separation ≥ 4.5 months, hemisphere R̄ ≥ 0.5, margin ≥ 1.0 month, exporter seasonality gate swept. Bottom table: gate 0.25 and margin 1.0 held, the clock’s own two parameters swept. The share outside HS 01-24 is the false-positive measure because no manufactured good has a harvest clock; it is a floor, not the full error rate, since an agricultural false positive is invisible to it.

Parameter sweeps: data/parquet/instruments_counterpoint_sweep_margin.parquet and data/parquet/instruments_counterpoint_sweep_latclock.parquet.

Query

SELECT strength_gate, n_products, n_lines, n_flagged, flagged_share_of_lines,
       flagged_share_outside_agri, NLD_flagged_value_share, HKG_flagged_value_share,
       RUS_flagged_value_share, CHN_flagged_value_share
FROM read_parquet('data/parquet/instruments_counterpoint_sweep_margin.parquet')
WHERE margin_months = 1.0
ORDER BY strength_gate

/* the clock's own two parameters */

SELECT sep_min, hemi_rbar_min, n_products, n_lines, n_flagged,
       flagged_share_outside_agri, flagged_value_share_global
FROM read_parquet('data/parquet/instruments_counterpoint_sweep_latclock.parquet')
WHERE (sep_min = 3.0 AND hemi_rbar_min = 0.3)
   OR (sep_min = 3.5 AND hemi_rbar_min = 0.4)
   OR (sep_min = 4.0 AND hemi_rbar_min = 0.5)
   OR (sep_min = 4.5 AND hemi_rbar_min = 0.5)
   OR (sep_min = 4.5 AND hemi_rbar_min = 0.6)
   OR (sep_min = 5.0 AND hemi_rbar_min = 0.6)
   OR (sep_min = 5.0 AND hemi_rbar_min = 0.7)
ORDER BY sep_min, hemi_rbar_min

Year	Fixed panel reporters	Unbalanced reporters	Within-series R̄, fixed	Within-series R̄, unbalanced	Overstatement	World resultant, fixed
2010	61	115	0.0930	0.0997	+7.1%	0.0299
2011	61	119	0.0915	0.0996	+8.8%	0.0202
2012	61	120	0.0896	0.1204	+34.4%	0.0052
2013	61	118	0.0907	0.0936	+3.2%	0.0119
2014	61	121	0.0870	0.0929	+6.8%	0.0134
2015	61	138	0.0876	0.0891	+1.7%	0.0101
2016	61	137	0.0873	0.0901	+3.2%	0.0169
2017	61	135	0.0828	0.0937	+13.1%	0.0192
2018	61	137	0.0832	0.0869	+4.4%	0.0079
2019	61	130	0.0788	0.0838	+6.4%	0.0059
2020	61	125	0.1277	0.1386	+8.5%	0.0648
2021	61	113	0.0957	0.1018	+6.4%	0.0158
2022	61	106	0.1030	0.1051	+2.1%	0.0258

Monthly reporter coverage jumps from 38 in 2009 to 117 in 2010, peaks at 140 in 2015 and reads 76 in 2023 only because the clean segment stops in August. MEASURED Every pooled statistic on this page is therefore computed on a fixed panel of 61 reporters that filed at least 11 months in every year from 2010 to 2022. It matters: the unbalanced panel overstates value-weighted within-series monthly concentration in all 13 years, by up to 34.4% in 2012. The aggregate world resultant, by contrast, is near zero on every panel, 0.0052 in 2012 to 0.0648 in 2020. Counterpoint cancels in the sum, which is exactly why the within-series statistic replaced it.

The 2009-to-2010 jump is 38 to 117 counting all reporter codes, or 38 to 115 after dropping rows with a NULL reporter ISO3. Both are shown because they are different numbers and the difference is the aggregate reporters. The world resultant column is the resultant of aggregate world monthly export value, a single series per year, and it is too unstable to headline; it is published so nobody has to take that judgement on trust.

Query

SELECT year, reporters_all_codes, reporters_iso, rows, value_usd, reporters_fixed_panel
FROM read_parquet('data/parquet/instruments_counterpoint_coverage.parquet')
ORDER BY year

/* fixed panel against the unbalanced one */

SELECT year,
       max(reporters) FILTER (WHERE panel = 'fixed_61') AS fixed_reporters,
       max(reporters) FILTER (WHERE panel = 'unbalanced_all') AS unb_reporters,
       max(vw_within_Rbar) FILTER (WHERE panel = 'fixed_61') AS fixed_within,
       max(vw_within_Rbar) FILTER (WHERE panel = 'unbalanced_all') AS unb_within,
       max(vw_within_Rbar) FILTER (WHERE panel = 'unbalanced_all')
         / max(vw_within_Rbar) FILTER (WHERE panel = 'fixed_61') - 1 AS overstate,
       max(world_Rbar) FILTER (WHERE panel = 'fixed_61') AS fixed_world_Rbar,
       max(world_Rbar) FILTER (WHERE panel = 'unbalanced_all') AS unb_world_Rbar
FROM read_parquet('data/parquet/instruments_counterpoint_panel.parquet')
GROUP BY year
ORDER BY year

3.0	0.3	594	4,322	319	41.7%	5.26%
3.5	0.4	424	3,394	214	38.3%	3.17%
4.0	0.5	268	2,292	126	34.9%	2.49%
4.5	0.5	193	1,676	99	33.3%	3.25%
4.5	0.6	155	1,363	50	18.0%	1.97%
5.0	0.6	99	859	36	8.3%	2.24%
5.0	0.7	68	661	23	4.3%	1.38%