Fetching primary parquet sources and computing exhibits.
research, network centrality
Which countries are structurally central in the global production network?
Treating the world economy as a weighted graph of intermediate-goods trade, a small set of countries anchor the network: every large flow either starts at them, passes through them, or ends at them. The question is not 'who exports most', that is already answered by trade tables , but who is most wired into the set of other well-wired countries. Two standard tools from network analysis make this precise: eigenvector centrality (Bonacich 1972; Newman 2010, chapter 7) and weighted-degree strength (Barrat, Barthélemy, Pastor-Satorras & Vespignani 2004, PNAS). When the graph is the production network, the centrality ranking is the map of whose disruption would propagate furthest into aggregate output (Acemoglu, Carvalho, Ozdaglar & Tahbaz-Salehi 2012, Econometrica).
Method and data caveat. We proxy intermediate goods by HS chapters 84 (mechanical machinery and parts) and 85 (electrical machinery and parts), which capture the bulk of BEC categories 41, 42, 51, and 53 (parts and capital goods) in the UN BEC Rev.5 - HS correspondence (UNSD, 2016). The workbench parquet build exports BACI bilateral flows only as total value, not per HS chapter, so we weight each directed bilateral flow fij by the exporter's own HS 84-85 share of its total exports in the same year. That is a coarser approximation than a true per-HS bilateral network (which would require reprocessing the raw BACI CSV), but it preserves cross-country ordering when HS 84-85 intensity varies systematically by exporter, as it does here. Eigenvector centrality is computed by power iteration on the symmetric weighted adjacency Wij = intensityi·fij + intensityj·fji.
latest year2024
baseline2000
nodes (top-50)50
directed edges (2024)2,433
most central (2024)CHN
most central (2000)USA
Who sits at the centre in 2024
Eigenvector centrality answers a recursive question: a country is central if it trades heavily with other central countries. The leading eigenvector of W encodes that definition uniquely under Perron-Frobenius (Newman 2010, §7.2). We rescale to [0, 1] by dividing through by the maximum; the top-ranked country scores 1.000. Countries not inside the top 50 by intermediate-proxy flow are excluded from the eigenvector computation.
Figure 1
Top 20 countries by eigenvector centrality in the intermediate-goods trade network, 2024
China (CHN) leads at 1.000, followed by USA (0.894) and MEX (0.455). The pattern is a narrow core of machinery-and-electronics hubs sitting at the top, tapering into a long tail of still-connected but less-wired exporters. This is the 'disassortative core' that Fagiolo, Reyes & Schiavo (2010) document for the world trade web: big traders link preferentially to other big traders.
Source: CEPII BACI bilateral trade 2024, this workbench parquet build. HS 84-85 intensity from country_year_product + country_year_product_ext (latest revision per country). Method: Bonacich (1972) eigenvector centrality on intensity-weighted symmetric adjacency; Newman (2010, OUP) §7.2. Authors calcs.
Who rose and who fell since 2000
A static centrality ranking is informative; the change in rank across a quarter-century is where the structural story sits. For every country in the top-50 panels of both 2000 and 2024, we take its eigenvector rank in each year and report the difference (rank2000 − rank2024). A positive value means the country has moved toward the core of the network; a negative value means it has fallen back. Carvalho (2014) emphasises that such shifts typically reflect a reorganisation of global value chains rather than a pure scale-up of trade.
Figure 2
Top 20 countries by eigenvector-centrality rank change, 2000-2024
Is centrality just size, or is something else going on?
The eigenvector ranking above has an obvious competing explanation: countries are central because they are big, and GDP already tracks economic mass. The production-network literature argues otherwise. Acemoglu et al. (2012) show that aggregate volatility depends on the eigenvector centrality of the input-output network, not on GDP per se, and that concentration of centrality in a few sectors or countries is what translates idiosyncratic shocks into macro fluctuations. An equivalent cross-country prediction: country-level centrality and GDP should be positively but imperfectly correlated, with identifiable outliers where centrality exceeds what GDP alone would suggest (classic trade hubs: Singapore, Netherlands, Belgium) or undershoots it (large commodity exporters).
Figure 3
Country eigenvector centrality versus GDP, 2024, log-log
Toy shock: if one country's exports are halved, who feels it first?
Acemoglu et al. (2012) show formally that a negative productivity shock at a central node propagates through the input-output graph with a magnitude proportional to that node's centrality. Here we run a much simpler, first-order version: halve country X's HS 84-85-proxy exports to every partner, compute the loss each partner takes as a share of its own intermediate-goods imports, and list the ten most exposed partners. We do this for three candidate shocks, China, the United States, and Germany, chosen because they are the three most central nodes in Figure 1. This captures only the direct bilateral hit: it ignores substitution, re-exports, and higher-order Leontief propagation, and is therefore a strict lower bound on true exposure.
Figure 4a
If CHN halved its HS 84-85-proxy exports, top 10 partners by loss as share of their intermediate-goods imports, 2024
Figure 4b
If USA halved its HS 84-85-proxy exports, top 10 partners by loss as share of their intermediate-goods imports, 2024
Figure 4c
If DEU halved its HS 84-85-proxy exports, top 10 partners by loss as share of their intermediate-goods imports, 2024
How large is this slice of trade, historically?
HS chapters 84 and 85 together have been the largest single block of world merchandise trade for decades. The time series below reports their combined share of world exports in the BACI parquet, 1995-2024, as context for the centrality numbers above: this is a large enough subuniverse that rankings on it are not a statistical artifact of a thin aggregate.
Figure 5
HS 84-85 share of world merchandise exports, 1995-2024
In 1995 the HS 84-85 share of world exports was 27.8%; in 2024 it was 27.3%. The level has been between 25% and 32% for the entire sample, confirming that the centrality network we analyse is the principal machinery-and-electronics backbone of world production.
Source: CEPII BACI via country_year_product, 1995-2024. Method: Σ(export_value where HS starts with 84 or 85) / Σ(export_value). Authors calcs.
Is the production-network core getting narrower?
Figure 1 is a static ranking; but centrality inequality across the top-50hub set is itself a time-varying object. We recompute eigenvector centrality on the intensity-weighted adjacency for eight years across the sample and report the Gini coefficient of the centrality vector in each year. Rising Gini = a shrinking, more top-heavy core. Falling Gini = the network is becoming more evenly wired. De Benedictis & Tajoli (2011, World Economy 34(11): 1417-1454) argued on earlier data that the world trade web had been concentrating; this is the eigenvector-centrality analogue.
Figure 6
Gini coefficient of eigenvector centrality across the top-50 intermediate-goods hubs, 1996-2024
In 1996 the Gini of centrality across the top-50 hub set was 0.636; in 2024 it was 0.593, a change of -0.042. The core has dispersed: centrality is more evenly distributed across the hub set in recent years, consistent with the nearshoring and diversification-by-design literature (Alfaro & Chor 2023).
Source: CEPII BACI bilateral flows and HS 84-85 intensity, 1996-2024. Method: for each year, recompute eigenvector centrality on the intensity-weighted top-50 adjacency, then compute the Gini coefficient (Gini 1912) on the centrality vector. Rising Gini = more top-heavy core. See De Benedictis & Tajoli (2011) World Economy 34:1417-1454 for the trade-web concentration motivation. Authors calcs.
Who rose and who fell in the post-GFC decade, 2010-2024
Figure 2 reports the long-run 2000-2024 shift; the post-global-financial-crisis decade is a distinct regime. Between 2010 and 2024the production network absorbed the China slowdown (Autor, Dorn & Hanson 2021, NBER WP 29401), the 2018-2020 tariff escalation (Fajgelbaum & Khandelwal 2022, ARE), COVID supply-chain disruptions, and the post-2020 nearshoring wave. This figure isolates that decade by reporting eigenvector-centrality rank change over the shorter window.
Figure 7
Top 20 countries by eigenvector-centrality rank change, 2010-2024
Which hubs have been most positionally volatile across regimes?
A stable central hub (DEU, USA, CHN) should look positionally similar across 2000, 2010, and 2024: the three snapshots span the China-shock, post-GFC, and post-COVID regimes, and a structurally-anchored hub should survive all three. We measure that by the standard deviation of rescaled eigenvector centrality across the three panels for every country present in all three with centrality above 0.10 in at least one. High sigma flags hubs whose production-network standing has oscillated regime-to-regime, typically the candidate 'switchers' whom Alfaro & Chor (2023) and Carvalho (2014) identify as GVC reshufflers.
Figure 8
Top 15 countries by three-panel centrality volatility (SD across 2000, 2010, 2024)
The spine: which bilateral edges carry the network
Figures 1 to 8 collapse the network into per-country summaries. Figure 9 reads the network the other way around and lists the single largest weighted bilateral edges in 2024: each entry is an exporter's flow to one specific importer, weighted by that exporter's HS 84-85 intensity, the same construction that feeds the eigenvector ranking. Hidalgo, Klinger, Barabási & Hausmann (2007, Science 317: 482-487) make the case for inspecting the underlying edges directly: the top-of-rank hubs in Figure 1 should be those that own the heaviest pairwise links, and the spine view validates or refutes that.
Figure 9
Top-15 weighted bilateral edges in the HS 84-85 production network, 2024
What this adds
The core is narrow and has tilted east. Figure 1 shows a handful of countries holding eigenvector centrality above 0.5 on the rescaled index; Figure 2 shows that the movement from 2000 to 2024is largely a rotation of that core toward Asia and toward near-shoring beneficiaries, consistent with the qualitative evidence in Alfaro & Chor (2023).
Centrality is not just GDP. The log-log correlation in Figure 3 is high but not unity, and the residual structure isolates the classic trade hubs (Singapore, Netherlands, Belgium) from the non-hub-large-economy profile (commodity exporters). This cross-section is the country-level counterpart of the sectoral residual structure Carvalho (2014) identifies for the US input-output table.
Exposure is concentrated on near neighbours, not biggest partners.Figure 4 shows that the import-side hit from halving a hub's HS 84-85 exports falls disproportionately on geographic and block-partner neighbours, not on the largest partner by absolute value. This is the first-order mechanism behind the 'trade-war contagion' literature (Fajgelbaum & Khandelwal 2022, Annual Review of Economics).
Open questions
Eigenvector centrality here uses a symmetric intensity-weighted adjacency; Katz centrality with a decay parameter α (Newman 2010, §7.4) would let us trade off direct vs. indirect exposure and is the natural robustness check. So would betweenness centrality on a sparsified top-k-edges version of the graph.
A true per-HS bilateral network, BACI HS6 × exporter × importer, would replace the exporter-intensity approximation and let us study machinery vs. electronics separately. That requires reprocessing the raw BACI CSV into a workbench-side partitioned parquet; out of scope here.
The first-order shock in Figure 4 is a strict lower bound on exposure. A Leontief inversion on a world IO matrix (OECD ICIO, Eora) would propagate shocks through all higher-order input linkages per Acemoglu et al. (2012, Prop. 2).
References
Acemoglu, D., Carvalho, V. M., Ozdaglar, A., & Tahbaz-Salehi, A. (2012). 'The Network Origins of Aggregate Fluctuations.' Econometrica 80(5): 1977-2016.
Alfaro, L., & Chor, D. (2023). 'Global Supply Chains: The Looming 'Great Reallocation'.' NBER Working Paper 31661.
Autor, D. H., Dorn, D., & Hanson, G. H. (2013). 'The China Syndrome: Local Labor Market Effects of Import Competition in the United States.' American Economic Review 103(6): 2121-2168.
Barrat, A., Barthélemy, M., Pastor-Satorras, R., & Vespignani, A. (2004). 'The architecture of complex weighted networks.' Proceedings of the National Academy of Sciences 101(11): 3747-3752.
Bonacich, P. (1972). 'Factoring and weighting approaches to status scores and clique identification.' Journal of Mathematical Sociology 2(1): 113-120.
Carvalho, V. M. (2014). 'From Micro to Macro via Production Networks.' Journal of Economic Perspectives 28(4): 23-48.
Fagiolo, G., Reyes, J., & Schiavo, S. (2010). 'The evolution of the world trade web: a weighted-network analysis.' Journal of Evolutionary Economics 20(4): 479-514.
Fajgelbaum, P. D., & Khandelwal, A. K. (2022). 'The Economic Impacts of the US-China Trade War.' Annual Review of Economics 14: 205-228.
Hidalgo, C. A., Klinger, B., Barabási, A.-L., & Hausmann, R. (2007). 'The Product Space Conditions the Development of Nations.' Science 317(5837): 482-487.
Newman, M. E. J. (2010). Networks: An Introduction. Oxford University Press, chapter 7.
United Nations Statistics Division (2016). Classification by Broad Economic Categories, Revision 5. UNSD Series M, No. 53, Rev.5, Annex III (BEC - HS correspondence).
The biggest riser is VNM (Viet Nam), up 36 positions (rank 43 in 2000 → rank 7 in 2024). The biggest faller is FIN (Finland), down 15 positions (rank 29 → rank 44). The composition tracks the well-documented GVC reshuffling after China's WTO accession (Autor, Dorn & Hanson 2013, AER) and the post-2018 nearshoring wave (Alfaro & Chor 2023, NBER 31661).
Source: eigenvector-centrality ranks for 2000 and 2024, computed as in Figure 1. Positive bars are rank gains (movement toward centre), negative bars are rank losses. Authors calcs.
The log-log correlation is 0.77 across the 49 countries with GDP data, with slope 0.78. Countries well above the regression line are 'over-central' relative to GDP: in 2024 that set is led by VNM, HKG, MYS, SGP, THA. Countries well below the line are 'under-central', big economies whose trade is less wired into the machinery-and-electronics backbone: GRC, NOR, PRT, IRQ, EGY. This separation is the empirical signature of the production-network channel (Carvalho 2014): the ranking reflects who is plugged into the intermediate-goods fabric, not just who is large.
Sources: eigenvector centrality from Figure 1; GDP from World Bank WDI NY.GDP.MKTP.CD, 2024 (current USD). Log-log OLS fit. Authors calcs.
Most-exposed partner: PRK (49.8% of its HS 84-85-proxy imports, $430.7M absolute loss). Total intermediate-proxy flow at stake in this toy scenario: $851.4B. The exposure ranking is dominated by geographic neighbours and block partners: a CHN disruption shows up first where the bilateral is deepest, not where the partner is largest.
Most-exposed partner: SXM (40.6% of its HS 84-85-proxy imports, $83.0M absolute loss). Total intermediate-proxy flow at stake in this toy scenario: $216.7B. The exposure ranking is dominated by geographic neighbours and block partners: a USA disruption shows up first where the bilateral is deepest, not where the partner is largest.
Most-exposed partner: AUT (17.0% of its HS 84-85-proxy imports, $8.9B absolute loss). Total intermediate-proxy flow at stake in this toy scenario: $201.1B. The exposure ranking is dominated by geographic neighbours and block partners: a DEU disruption shows up first where the bilateral is deepest, not where the partner is largest.
In the post-GFC decade, the biggest riser is VNM (Viet Nam), up 20 positions (rank 27 in 2010 → rank 7 in 2024). The biggest faller is BEL (Belgium), down 7 positions (rank 24 → rank 31). Countries that move sharply here but little in Figure 2 are post-GFC structural shifters rather than post-WTO ones; the contrast between the two windows separates the China-shock era from the reshoring era.
Source: eigenvector-centrality ranks for 2010 and 2024, computed as in Figure 1 on the intensity-weighted top-50 adjacency. Positive bars = rank gains; negative = losses. Authors calcs.
The most positionally volatile hub on this measure is CHN (China), with centrality moving 1.00 → 0.37 → 1.00 across 2000, 2010, and 2024 (SD = 0.298). Low-volatility hubs (DEU, USA) anchor the production network across regimes; high-volatility hubs have moved through the core rather than sitting in it, and are the natural candidates for the reshoring/nearshoring rearrangement literature (Alfaro & Chor 2023, NBER 31661; Carvalho 2014, JEP).
Source: CEPII BACI bilateral flows and HS 84-85 intensity, 2000, 2010, and 2024. Method: for each country appearing in the top-50 panel in all three years with centrality >= 0.10 in at least one, compute the standard deviation of rescaled eigenvector centrality across the three snapshots. High SD flags hubs whose production-network standing oscillated across the China-shock/post-GFC/post-COVID regimes. Authors calcs.
The largest single edge in 2024 is CHN → USA at $212.8B of intensity-weighted intermediate-goods flow. The top-15 edges together carry roughly 22.0% of the total intensity-weighted flow across the entire bilateral graph in 2024. The list is dominated by exports out of the Figure 1 leaders into the largest importing markets, which is the cross-check the page promises: a centrality ranking that did not coincide with the heaviest single edges would be a measurement artefact. The asymmetric structure of this list is the bilateral analogue of the asymmetric Leontief propagation in Acemoglu, Carvalho, Ozdaglar & Tahbaz-Salehi (2012, Econometrica): a sectoral or country shock at the source of any of these edges would propagate disproportionately to its single named partner.
Source: CEPII BACI 202501 (retrieved 2026-04-28) bilateral_year, 2024, restricted to bilateral pairs with both endpoints in the top-50 HS 84-85 trader set; flows weighted by exporter HS 84-85 intensity (same construction as Figure 1 / Newman 2010 §7.2). Sorted by weighted flow, top 15 reported. Reference: Hidalgo, Klinger, Barabási & Hausmann (2007, Science 317).