What does the world trade network look like?

Treat countries as nodes and bilateral trade flows as weighted edges, and the world economy becomes a graph. A few hubs anchor the structure; regional clusters form; the whole object tightens over time. This page recomputes the basic network statistics economists use when they study trade as a complex network, following De Benedictis & Tajoli ( World Economy, 2011) and Fagiolo, Reyes & Schiavo ( Journal of Evolutionary Economics, 2010).

Method note. We take the Top-50 countries by total trade in a given year, build the undirected weighted graph whose edge W_ij = Xᵢⱼ + Xⱼᵢ, and run (i) weighted-degree centrality (Barrat et al. 2004 PNAS), (ii) eigenvector centrality by power iteration on that same W, (iii) label-propagation community detection (Raghavan, Albert & Kumara 2007 Phys. Rev. E) capped at 7 communities, and (iv) the global clustering coefficient C = 3·triangles / connected-triples on the Top-200 edges each year (Watts & Strogatz 1998; Newman 2003 §3.3). Label propagation is a Louvain-style proxy, not modularity optimization , communities here should be read as coarse regional clusters, not statistically tested partitions.

A few hubs hold it together

The classic finding in the trade-network literature is that bilateral flows are dramatically concentrated on a handful of hubs. Fagiolo, Reyes & Schiavo (2010) report that the world trade web is disassortative , big traders preferentially link to other big traders and to many small ones , and that it has a near scale-free weighted-degree distribution. The figure below draws the Top-100 undirected pairs in the Top-50 subgraph for 2022. Outer ring: ranks 1-15; middle: 16-30; inner: 31-50. Node area is proportional to the country's total trade (exports + imports).

Who is most central?

Two standard notions. Weighted degree centrality is just the sum of incident trade flows , a measure of raw connectedness. Barrat, Barthélemy, Pastor-Satorras & Vespignani (2004 PNAS) proposed this extension of classic degree centrality to weighted graphs. Eigenvector centrality measures how much a country trades with other well-connected countries , recursively (Bonacich 1972). We compute it by power iteration on the Top-50 weighted adjacency matrix. A country ranks high on eigenvector centrality only if it is well-connected to other hubs, so the two rankings can differ sharply.

Regional communities

The trade graph is not homogeneous: distance, language, RTAs and colonial history cluster countries. De Benedictis & Tajoli (2011) run community detection on the world trade network and find intuitive regional blocks , a North-American cluster, a European core, an East-Asian cluster , plus strong colonial-legacy communities in francophone Africa and the British Commonwealth. We run label propagation (Raghavan, Albert & Kumara 2007), which is an O(E) proxy for the Louvain modularity algorithm (Blondel et al. 2008), and cap the output at 7 communities. Same circular layout as Figure 1; color is community.

Is trade becoming more clustered?

The global clustering coefficient C asks: if A trades with B and B trades with C, how often does A also trade with C? Watts & Strogatz (1998) showed that real networks are much more clustered than random graphs of the same density. Fagiolo, Reyes & Schiavo (2010) track C on the world trade web and find it is stable and high across 1981-2000. We recompute on the Top-50 binary subgraph , keeping the Top-200 undirected edges each year to hold density roughly constant , and check whether clustering has moved over the BACI window.

Is the block structure getting sharper?

Figure 3 showed that label propagation recovers regional blocks. Whether those blocks are 'real' , i.e., whether the within-block edge weight exceeds what a strength-preserving null model would predict , is a separate question. Newman's (2004, Phys. Rev. E 69: 066133) modularity Q answers it. On an undirected weighted graph, Q = (1/2m) · Σ_ij [W_ij − k_ik_j/(2m)] · δ(c_i, c_j). Values above roughly 0.3 are conventionally read as meaningful community structure (Newman 2006, PNAS 103(23): 8577-8582); 0 is the configuration-model null. For 2022, the label-propagation partition reaches Q = 0.000 on the Top-50 weighted graph.

Which nodes hold the network together?

Centrality says who is big or well-connected; it does not say who is structurally load-bearing. Albert, Jeong & Barabási (2000, Nature 406: 378-382) propose a node-removal robustness test: delete a node, recompute a global structural statistic, report the drop. Big drops flag nodes the network 'needs' to keep its topology together. Following Serrano & Boguñá (2003, Phys. Rev. E 68: 015101) and Saramäki et al. (2007) on trade-network sensitivity, we use the global clustering coefficient as the target statistic: remove each Top-50 country in turn, recompute C on the remaining binary subgraph (Top-200 edges rule held constant), and report the five nodes whose removal drops C the most. These are the triangle-makers: pull one, and many connected triples lose their third side.

Is clustering uniform across node tiers?

The global clustering coefficient in Figure 4 is a single network-wide number. It hides whether the hub nodes (high weighted-degree) close their triangles at the same rate as the peripheral ones. Soffer & Vázquez (2005, Phys. Rev. E 71: 057101) and Saramäki et al. (2007, Phys. Rev. E 75: 027105) document a common 'clustering hierarchy' in real networks, where the local clustering coefficient C_v = 2·t_v / (k_v(k_v − 1)) tends to fall with degree , hubs are connectors, not closers. We bucket the Top-50 nodes by weighted-degree quintile (Q1 = periphery, Q5 = top traders) on the same Top-200-edge binary subgraph used in Figure 4, and report the mean local clustering within each bucket.

How heavy is the tail of the degree distribution?

Fagiolo, Reyes & Schiavo (2010) report that the world trade web has a near scale-free weighted-degree distribution. The cleanest visual test is the rank-size plot: order nodes by weighted degree, plot rank on a log axis against degree on a log axis. A straight line on log-log axes is the rank-frequency signature of a power law (Gabaix 2009, Annual Review of Economics 1: 255-294, on Zipf-style distributions and Newman 2005, Contemporary Physics 46(5): 323-351, on the rank-size diagnostic). Curvature near the head means the top traders are dominating even more than a pure power law would predict; curvature in the tail means the periphery is thinner than a power law allows.

What this tells us

Trade looks like a classic hub-and-spoke weighted graph. Four operational facts. (1) The top five countries (CHN, USA, DEU, JPN, KOR) account for 39% of total trade among the Top-50 in 2022; the network is disassortative in the sense of Fagiolo, Reyes & Schiavo (2010), with big traders preferentially linking to other big traders. (2) Eigenvector and weighted-degree rankings track each other at the top but can differ by 3+ positions in the middle, marking countries whose trade is concentrated on hubs rather than spread across the periphery. (3) Label propagation recovers roughly the regional blocks De Benedictis & Tajoli (2011) find , North American, European, East Asian , plus colonial-legacy clusters. (4) Clustering on a density-controlled Top-50 subgraph moved from 0.442 in 1995 to 0.404 in 2020(-0.038); a high and stable C is the network-analytic shadow of the gravity equation (Tinbergen 1962; Head & Mayer 2014): proximity and triangular closure drive flows, and globalisation has not eliminated either.

What this page is not: we do not estimate modularity, do not report significance of communities against a null model (e.g., the directed configuration model of Squartini & Garlaschelli 2011), and do not run Louvain proper. Label propagation is a proxy. The clustering coefficient we report is binary; weighted versions (Onnela et al. 2005, Saramäki et al. 2007) give different numbers and would be a worthwhile follow-up.