How dependent is a country on a few suppliers for a given product, how is that changing, and who could plausibly replace a supplier?
TradeGraph AI is an open, research-oriented trade-dependency explorer built on reconciled bilateral trade data. The Atlas of Economic Complexity, OEC and WITS already map trade flows; TradeGraph focuses on dependency. Every number carries its source release, denominator, units and coverage. Everything runs in the browser: the pipeline publishes Parquet files, and the app queries them with DuckDB-WASM, with no backend.
More in the gallery.
| Page | What it answers |
|---|---|
| Overview | Headline numbers computed live from the data, and entry points to every page. |
| Dependency profile | Who supplies country X with product P? How concentrated is that (shares, HHI, effective suppliers), by value and by tonnes, from 2012 to 2024? |
| Biggest movers | Where did concentration rise or fall most between 2012–14 and 2021–23? |
| EU monthly monitor | Month-by-month supplier shares for the EU and six member states (Eurostat); a nowcast of the years BACI has not reached yet; how imports cross the EU border (sea, air, road, rail). |
| Trade map | Where do flows run? Animated flows on a map or globe, coloured by importer. |
| Sea routes | Modelled sea lanes and chokepoints (Suez, Hormuz, Malacca, …): which importers rely on which passage, what a closure would reroute, and observed ship traffic. |
| Network | Which trade communities exist, who sits at their centre, and who is more central than a gravity model predicts? |
| Scenario studio | If a major supplier falls short, how much do others make up, under fixed or learned capacity, how much stays unmet, and how uncertain is that? (Labelled scenario.) |
| Findings, Forecast lab, Methodology | Written findings, a forecast benchmark against "same as last year", and how everything is computed. |
Findings:
- Rare earths and magnets: China's share of German and US imports rose from about 52–58% to 73–77% by value, driven by permanent magnets.
- Forecast benchmark: machine learning beats persistence by about 5% in log error, but last year's value remains very hard to beat for the largest flows.
Boundary: aggregate trade data are not firm supply chains, sea routes are modelled rather than tracked, and scenarios are assumptions, not forecasts or causal estimates.
| Source | Use | Licence |
|---|---|---|
| CEPII BACI HS12 V202601 | All trade flows (HS6, ~200 countries, 2012–2024; value and tonnes) | Etalab Open Licence 2.0 |
| CEPII Gravity V202211 | Distance, border, language, colonial ties, trade agreements | Etalab Open Licence 2.0 |
| Eurostat Comext DS-045409 | Monthly EU imports (HS6), 2017 to the latest month | Eurostat reuse policy (acknowledgement; adapted) |
| Eurostat Comext, trade by mode of transport | How EU imports from outside the EU cross the border, 2012–2026 | Eurostat reuse policy (acknowledgement; adapted) |
Eurostat ert_bil_eur_m |
Monthly euro to US dollar rates (ECB statistics), used only to convert values | Eurostat copyright notice; ECB statistics reuse policy (free reuse with credit) |
| IMF PortWatch | Ports, chokepoints and daily chokepoint transits | IMF terms: reuse with attribution; commercial reuse needs IMF permission |
| World Bank WDI | GDP (reserved) | CC BY 4.0 |
The project is non-commercial, which some of these terms require. All data is adapted (aggregated, converted, modelled), and the sources are not responsible for the adapted figures. Full attributions, citations and third-party notices are in DATA_LICENSES.md.
Product baskets, editable in configs/baskets.yaml: chips, Li-ion batteries, lithium chemicals, rare earths and magnets, bulk antibiotics, and green coffee as a control.
Details are in the data card and the model card.
Just the app. The published data the app reads is committed in app/public/data, so Node.js 20+ is enough:
cd app && npm ci && npm run dev # http://localhost:5173The full pipeline. Rebuilding the data from the raw sources needs uv (it installs Python 3.12+ itself) and about 7 GB of free disk:
uv sync # Python environment
uv run pytest # offline tests on tiny fixtures
uv run tradegraph all # downloads the sources, validates them and builds every mart (several minutes)
make dev # copies the new marts into the app and starts itIndividual steps: uv run tradegraph fetch | curate | marts | wdi | gravity | forecast | scenarios | comext | transport | sea | substitution | nowcast.
Different product baskets: edit configs/baskets.yaml, then run uv run tradegraph marts (seconds; nothing is re-downloaded), followed by the later steps. Years, releases and benchmark windows are set in configs/pipeline.yaml.
The app is a static site; the Python pipeline runs locally and is not deployed. On Vercel, import the repository and set Root Directory to app. app/vercel.json does the rest (Vite build, output dist, caching). Any static host works the same way: run cd app && npm ci && npm run build and serve app/dist.
raw snapshots (SHA-256 manifests; git-ignored)
─► validated Parquet (DuckDB) ─► marts: dependency · movers · gravity/network · forecast ·
scenarios · Comext · nowcast · transport mode · sea routes
─► app/public/data ─► React + Vite app (deck.gl, Observable Plot, DuckDB-WASM in the browser)
docs/data_card.md·docs/model_card.md·docs/gallery.mdsite/: the first prototype (Observable Framework), kept for reference; the app inapp/replaces it
The code is released under the MIT License. The data in app/public/data, reports/ and docs/screenshots is derived from the sources above and stays under their terms; see DATA_LICENSES.md.




