Pillars 1 + 2 · Talk vs trade · Methodology

How we raced talk against trade

We found lasting shifts in EU trade and in EU talk with the same tool, matched them by a rule we wrote down before looking, and then asked whether the matches beat chance.

Open the data folder (Google Drive) Back to Talk vs trade

01The question and the plan
02The two sides
03Finding lasting shifts
04Matching and reading lags
05The chance test
06Whole-series tests
07Limits and AI use

01 · The question and the plan

Does talk move before trade, after it, or not at all?

Every rule was frozen on 31 Oct 2026, before a single trade shift had been computed, so the answer could not steer the method.

Why a frozen plan matters

With many series, many shifts and a flexible idea of “close in time”, almost any story can be found after the fact. So we fixed in writing, with a date stamp: which talk series pairs with which trade series, the window, how a lag is computed, what counts as “talk first”, and the random seed (20261031).

Pairing by stage

Each talk series is paired with the trade line that sits at the same stage of the chain. The headline talk shift, “no stage named”, is not a stage, so it is paired with the most battery-specific trade line, anode-grade graphite.

Dropped from the plan: packs & vehicles talk against battery imports (CN 8507.60), and lithium and cobalt comparisons, because the trade file does not contain those series.

Window

Comparison window Feb 2022 – Jul 2026 (the end of the trade data). Trade shifts are fitted on Jan 2019 – Jul 2026 to give the tool a longer baseline, but only shifts from Feb 2022 on are compared.

ID	Talk series	Trade series	Why
P1	No stage named · all texts	Anode-grade natural graphite	The headline talk shift. Not a stage, so compared with the most battery-specific trade line.
P1n	No stage named · the press	Anode-grade natural graphite	The press carries the Oct 2025 rise.
P2a	Refining talk	Artificial graphite (CN 3801.10)	Synthetic graphite is a refined product.
P2b	Refining talk	Refined natural graphite (2504.10, €1,200–3,699/t)	Natural graphite that has been processed but is not yet anode grade.
P3	Components talk	Anode-grade natural graphite (2504.10, ≥ €3,700/t)	Anode material is a battery component.
P4	Mining talk	Natural graphite, all (2504.10)	Closest code to raw ore.
G1	Graphite only · no stage named (quarterly)	Anode-grade natural graphite (quarterly)	Graphite-specific check; thin.
G2	Graphite only · refining talk (quarterly)	Artificial graphite (quarterly)	Graphite-specific check; thin.

02 · The two sides

Trade is China’s share of what the EU imports; talk is the share of paragraphs making each kind of claim.

Both are monthly. Trade comes from Pillar 1; talk comes from Pillar 2.

Trade (Pillar 1)

Source: Eurostat Comext, EU27 imports from outside the EU, prepared by the Pillar 1 team.
Main measure: China’s share of import value, in percent: imports from China ÷ imports from all non-EU countries. Recomputed from values in euros. A month with no imports at all is filled in from its neighbours.
Second measure: volume, as the log of tonnes imported from China. Logs stop a few huge months from dominating.
Price bands. Natural graphite (CN 2504.10) mixes rough flake with near-anode material. Pillar 1 split it by unit value: €1,200–3,699 a tonne counts as refined, €3,700 or more as anode grade. One threshold was used.
Provisional months: May – Jul 2026.

Talk (Pillar 2)

Source: 14,577 paragraphs from EU institutions and the press, coded by Claude and adjusted for coding errors (see the Pillar 2 methodology).
Measure: each month, the share of all paragraphs that claim a dependency and place it at a given stage, or name no stage.
Streams: all texts for every series; the press added for “no stage named”, where its shift is confirmed.
Graphite only: monthly graphite talk is too thin (median six paragraphs), so graphite pairs use quarters, matched with quarterly trade to Q2 2026.

03 · Finding lasting shifts

The same shift-finding tool runs on both sides, with the same settings.

A lasting shift is a month when a series settles at a new average level and stays there for at least six months.

Two methods must agree

PELT finds the set of break points that best splits a series into flat segments, paying a penalty (2σ² ln n) for each extra break. σ is estimated robustly from month-to-month changes.
Bai–Perron fits the best set of up to five breaks by least squares and picks the number with the BIC.
A shift is confirmed when both put it within two months of each other. Shifts found by only one method are kept and labelled.
Each shift gets a 95% range for its date (Bai 1997). These ranges are wide: often one to three years.
Sensitivity: we halved and doubled the PELT penalty and flag which shifts survive.

Shift sizes are the difference between the average level before and after, in points of China’s share (or of the paragraph share for talk).

Trade series	Measure	Months	σ	Breaks (BIC)
Natural graphite, all	China’s share	91	8.4	3
Natural graphite, all	Volume	91	0.22	2
Refined band	China’s share	91	12.5	3
Refined band	Volume	91	1.08	1
Artificial graphite	China’s share	91	5.1	3
Artificial graphite	Volume	91	0.22	3
Anode band	China’s share	91	13.0	3
Anode band	Volume	91	0.80	2

Fitted on Jan 2019 – Jul 2026. σ is in points of China’s share, or in log tonnes for volume.

04 · Matching and reading lags

Each talk shift is matched to its nearest trade shift; the gap decides the reading.

Lag = month of the trade shift minus month of the talk shift. Ties go to the earlier trade shift.

Three months is the shortest gap we read as one side leading. Inside ±2 months the two moved together, which usually means both reacted to the same event. Beyond a year we call the shift unpartnered. We also note whether the two date ranges overlap. They almost always do, which on its own says the order is uncertain.

05 · The chance test

If random dates match as often, a match tells us nothing.

Trade series shift several times in 54 months, so a talk shift landing within a year of one is likely by luck alone.

How it works

Keep each pairing’s number of talk shifts and trade shifts.
Place them at random months, at least six months apart (the same rule the tool uses).
Count how many talk shifts have a trade shift within 12 months.
Repeat 5,000 times. The p-value is the share of random histories with at least as many matches as the real one.

The real total is 10 matches out of 13 talk shifts; random histories average about 10.7. No pairing has p below 0.30.

Real matches

10/13

Talk shifts with a trade shift within a year (China’s share).

Expected by chance

10.7

Average over 5,000 random histories.

06 · Whole-series tests

Correlating every month, not only the shifts, finds nothing either.

Shifts throw away most of the data. These tests use all of it.

Cross-correlation

Every lag from −12 to +12

We correlate month-to-month changes (first differences, so trends do not create fake links) at each lag. Positive lags mean talk moves first. Overlap Feb 2022 – Jul 2026, 54 months.

Significance

2,000 shifted copies

We slide the trade series round in time 2,000 times and record the largest correlation at any lag. The real maximum must beat 95% of them. A Holm correction then accounts for testing six pairings. None passes.

Granger, exploratory

Does one help predict the other?

Tests at lags of 1 to 6 months in both directions, smallest p kept. With 32 tests on 53 months and shifts inside the sample, small p-values are expected by chance. Reported, not used as findings.

07 · Limits and AI use

The method can rule out a clear lead; it cannot see the most recent months.

Read the absence of a pattern as “no evidence that talk leads trade”, not as proof that they are unrelated.

Limits

Blind spot. A shift needs six months at its new level, and trade ends in Jul 2026. No trade shift after about Jan 2026 can be found, so the missing trade response to China’s Oct 2025 controls (suspended Nov 2025, until 10 Nov 2026) is only partly informative.
The anode band is a price construction. Chinese prices sit near the €3,700 line, so some of the Mar 2025 fall is goods moving into the refined band. The headline uses the whole code (−15 points), not the band (−51).
Tier-level talk shifts are found by one method only; only the two “no stage named” shifts are confirmed by both.
Same-time moves (Dec 2023 – Mar 2024) are most simply read as both sides reacting to China’s graphite licences, not one driving the other.
Gaps in the trade file: no anode filter on artificial graphite, no battery, lithium or cobalt series; the new code 3801.10.10 has only seven months.
Short series. 54 months and 18 quarters leave little power for any test.

How we used AI

Talk data come from Claude’s coding of each paragraph (Pillar 2), checked against hand coding and adjusted for its errors.
Code. Claude wrote the matching, chance and correlation scripts from the frozen plan; we ran and checked them.
Decisions stayed with us: the pairings, the lag zones, the 12-month window and the choice to keep the graphite-only pairs were set by the team before the trade shifts were run.
This website was built with Claude from our results.