Data & Methods

What the network is built from, and the modelling choices behind it

This page is the single reference for the data source and the analytical decisions. The analysis notebooks assume it rather than restating it.

Source — CEPII BACI

BACI is CEPII’s harmonised bilateral trade database, reconstructed from UN Comtrade with the mirror-flow discrepancies between what an exporter declares and what the importer records reconciled into a single value per flow. It covers 200+ countries at HS-6 product level and is the standard source for academic trade-network work.

It was chosen over two alternatives. The UN Comtrade API needs a registered key and caps rows per call on the free tier, so a full bilateral matrix takes many paginated requests. Eurostat Comext is EU-reporter-centric, so flows between two non-EU partners are thin — that asymmetry would distort a network where the partners are nodes in their own right.

Each row of the raw file is one exporter × importer × product × year observation. Two column facts matter analytically:

• v — trade value in thousands of current USD. Everything in this project is value, never volume.
• q — quantity in metric tons, missing in roughly 2% of rows. It is not used anywhere.

Raw files are not committed. See data/README.md in the repository for the one-time download.

Scope — the node set

The network is the EU-27 plus 8 major trading partners (GBR, CHE, NOR, USA, CHN, TUR, RUS, JPN) = 35 economies, for the year 2022. Both the year and the country set live in src/eu_trade_network/config.py and can be changed without touching any notebook.

Products are aggregated over all HS chapters for the merchandise network, and restricted to HS chapter 27 (mineral fuels, oils, gas, electricity) for the energy subnetwork in RQ4. Self-loops are dropped, and a flow is kept only when both endpoints are in the node set — a restriction whose cost is measured explicitly in RQ4.

How the network is modelled

The graph is directed and weighted. An edge i→j carries value_kusd, the exports from i to j. So a node’s out-strength is its total exports and its in-strength its total imports. The result is a complete digraph: every economy trades with every other, giving 1,190 edges at density 1.0. That single property drives most of what follows.

Betweenness needs a distance, not a capacity

Trade value is a capacity — a bigger number means two economies are more strongly connected, i.e. closer. Shortest-path algorithms want the opposite: a cost, where bigger means further apart. Passing value_kusd straight into betweenness would therefore route every path through the weakest links and invert the result.

Weighted betweenness here uses distance = 1 / value_kusd, so heavily-traded corridors are short and thin ones are long. Unweighted betweenness is reported alongside it as the pure-topology view — and on a complete digraph it is identically zero, because every pair is already adjacent and no node ever lies on a shortest path.

Eigenvector centrality and PageRank use value_kusd directly as the weight, since those measures want a strength, not a distance.

The backbone, not a power law

Because the graph is dense, the scale-free and small-world framing that suits sparse networks does not apply, and no power law is fitted to the strength distribution. Where a sparse object is genuinely needed — for community detection and for asking whether the network can fragment — we extract the statistically significant backbone with the disparity filter (Serrano, Boguñá & Vespignani, PNAS 2009).

The filter asks, for each economy separately, whether a given link carries a significantly larger share of that economy’s trade than a random allocation of its total across its partners would produce. Links that survive at α = 0.05 are the ones that matter to at least one of their two endpoints. This keeps 184 of 1,190 edges — a network that is still connected but now has structure to find.

Communities

Louvain runs on an undirected, weighted projection: the i→j and j→i values are summed into one edge, so a bilateral relationship is a single weighted link. It is run on both the full graph and the backbone, and the partition with the higher modularity Q is kept.

Resilience

Economies are removed one at a time, tracking two quantities after each removal: the largest weakly-connected component as a share of surviving economies, and the share of the original trade value still carried inside that component. Value stranded in a splinter component counts as lost.

Targeted removal takes hubs first, by out-strength and then by weighted betweenness, with both rankings fixed on the intact network rather than recomputed as it degrades. Random failure is the mean of 20 seeded runs, shown with a ±1 sd band. The critical fraction is the first removal share at which the largest component stops holding half the surviving economies.

Reproducibility

Every stochastic step — Louvain, the random-failure runs, network layouts — is seeded from config.RANDOM_SEED, so re-running the notebooks reproduces the published numbers exactly.

Caveats worth reading before the results

2022 is not a neutral year. It is the energy-crisis year and the first year of large-scale sanctions on Russia. BACI values are current USD, so the energy subnetwork’s share of total value reflects prices as much as volumes, and Russia still tops the 2022 energy table despite sanctions that were only partly in force during it.

The node set is a frame, and for energy a tight one. Only 48% of these economies’ energy imports originate inside the node set, against 65% for merchandise. Europe’s crude and LNG largely come from the Gulf, Kazakhstan, and North and West Africa — none of which are nodes here. The energy graph therefore describes how energy is redistributed within Europe, not where it comes from.

Customs data misattributes pipeline gas. A flow is recorded against the declared consignment partner, so gas crossing trading hubs lands on the wrong edge. In this data Austria looks like a net gas exporter in 2022, which it certainly was not. Treat trade data as evidence about value routed, and use physical flow data for questions about dependence.

Licence and citation

BACI is distributed under the Etalab Open Licence 2.0 — any use is permitted with attribution, which is why this repository points at CEPII rather than redistributing the files.

Gaulier, G. & Zignago, S. (2010). BACI: International Trade Database at the Product-Level. CEPII Working Paper N°2010-23.