Inequality in economic shock exposures across the global firm-level supply network

The study addresses how economic shocks propagate through the global firm-level supply network and how countries are exposed to direct and indirect production losses from firm failures abroad. While globalization has boosted growth and reduced some dimensions of inequality, international production networks transmit shocks across borders. Prior analyses often use aggregated country or sector input–output data, potentially missing crucial firm-level topology that governs shock spreading. The authors develop a microscopic, network-based measure—generalizing DebtRank to supply networks—to quantify, for each country, expected exposure to losses when a random firm fails in another country. They distinguish downstream (supply) and upstream (demand) cascades, compare exposures with measures of gains from globalization (GDP per capita and growth), and investigate the global distribution and inequality of exposure risks, providing a new perspective on the benefits and risks of globalized production.

The paper builds on strands of research showing that international trade affects growth, inequality, and sustainability, but also externalizes environmental and labor risks. It draws from financial network systemic risk literature (e.g., DebtRank) where a node’s systemic importance depends on network position, not just size. Recent work has adapted such concepts to production networks and used agent-based or regional adaptive input–output models to assess disaster impacts. Empirical evidence from U.S. natural disasters and the 2011 Great East Japan Earthquake demonstrates substantial shock propagation along supply chains, including cross-border effects. Traditional country/sector-level IO approaches overlook firm-level topology. This work extends systemic risk quantification to global firm-level supply networks, emphasizing the necessity of granular network structure for understanding shock spreading and systemic exposures.

Data: Supplier–customer relationships were obtained from S&P Capital IQ (2017). Starting from 1,403,807 inter-firm relationships across 206 countries, the analysis restricts to supplier links (968,627) to capture flows of goods/services. Firms lacking location or sector information were removed; countries with ≤30 firms and known offshore jurisdictions (e.g., Cayman Islands) were excluded. The resulting unweighted, directed network contains N=230,970 firms and L=660,701 links after removing isolates, parallel links, and self-loops.

Shock propagation and exposure metrics: The authors adapt DebtRank’s cascading distress mechanism to supply networks. They simulate short-term cascades triggered by the complete failure of a single firm, allowing shocks to propagate through supplier–customer links until convergence, without recovery dynamics. Downstream cascades model supply shortages (supplier failure reduces customer output proportional to unavailable input shares); upstream cascades model demand reductions. The approach is agnostic to the shock cause and associates each firm with a worst-case systemic impact under these assumptions.

Definitions: Firm–Country Exposure E_id is the fraction of country d’s national production lost when firm i fails, aggregating direct and indirect effects via the network. Country–Country Exposure E_cd is the average exposure of country d to random firm failures in country c: E_cd = (Σ_{i in c} E_id)/C, where C is the number of firms in c. Absolute exposed value V_cd = k_d E_cd, where k_d is the economic size proxy of country d (sum of firm degrees). Firm size is proxied by degree k_i; country size by total degree k_c = Σ_{i in c} k_i. Exposures are computed separately for downstream (E_cd^down) and upstream (E_cd^up) cascades, with total exposure E_cd = E_cd^up + E_cd^down. Expected loss incorporating heterogeneous default probabilities p_i is defined as E = (Σ_i p_i E_id)/(Σ_i p_i), explored in supplementary analyses.

Empirical analyses: The authors characterize regional structure by sorting countries by continent/region and comparing E_cd^down to the country-level adjacency A_cd (number of links). They aggregate countries into low-, middle-, and high-income groups (by GDP per capita) to analyze asymmetric exposures between income groups. They fit gravity models to A_cd and E_cd^down, perform correlation analyses between exposure and GDP per capita, and run multivariate regressions (log–log) with GDP per capita as the dependent variable and controls (GDP, imports, exports). Inequality is quantified via Lorenz curves and Gini coefficients for exposure versus GDP. Robustness checks include: alternative normalization by total degree; simulations with simple input substitution; incorporating heterogeneous default probabilities; sampling experiments to assess dataset coverage bias; comparison with a VAT-based, weighted firm-level network from Ecuador (testing the effect of removing low-value links and ignoring weights).

• Country–country exposures (E_cd^down) show strong regional clustering and are highest within countries; block-diagonal patterns by continent/region emerge, far more pronounced than in the country adjacency A_cd.
• Exposures are asymmetric and non-reciprocal. Aggregating by income groups reveals: high-income countries, on average, impose greater downstream exposure on middle- and low-income countries than vice versa; countries are most exposed to firms within their own income group; middle-income countries bear the largest exposure overall.
• Total downstream exposure E_cd^down is significantly anti-correlated with GDP per capita (Pearson r = −0.52, p < 1e−15). A multivariate log–log regression with controls (GDP, imports, exports) finds E_cd^down has a significant negative association with GDP per capita (adjusted R^2 = 0.50, p < 1e−15), indicating poorer countries are more exposed.
• No evidence of a risk premium: downstream exposure shows little to no correlation with average GDP per capita growth over the past 20 years (r = 0.15, p = 0.12).
• Inequality of exposure is extreme: Lorenz analysis indicates the least exposed 80% of the global population accounts for ~10% of total exposure; the most exposed 20% account for ~90%. Gini coefficients: exposure 0.83 vs GDP 0.59.
• Upstream cascades show differing patterns: most exposure occurs between low- and middle-income countries; high-income countries neither create nor receive large upstream exposure. Upstream exposure correlates negatively with GDP per capita (r = −0.20, p < 0.04) with a Gini ~0.81.
• While V_cd correlates strongly with the average number of links k_cd (Pearson r = 0.93, p < 1e−15), exposure varies by up to two orders of magnitude for a given k_cd, underscoring the importance of detailed firm-level topology. Classic gravity models explain A_cd well but not E_cd^down; E_cd^down depends only weakly on exposure created by c and inversely on exposure received by d, suggesting exposures arise from many small links rather than few large ones.

The findings demonstrate that firm-level network structure critically shapes international shock propagation and that exposure to systemic production losses is highly uneven across regions and income levels. Countries are most exposed to firms within their own economies and nearby regions, but wealthy countries export systemic risk beyond their regions, disproportionately affecting poorer economies. Contrary to expectations, higher exposure is not compensated by higher GDP growth, challenging narratives of uniform net gains from globalization for developing countries. The pronounced inequality of exposure—greater than income inequality—reveals a new dimension of global disparity.

The results support the necessity of granular, firm-level data to assess systemic risk flows, as country-level link counts (A_cd) and gravity models cannot capture shock-spreading heterogeneity. Normalization checks suggest size effects do not drive the observed patterns. Policy implications include: building global infrastructures to map and monitor supply networks; restructuring production networks to distribute exposure more fairly (aligned with SDG 10); and adapting systemic risk management tools (e.g., a systemic risk tax) to supply chains to incentivize resilience.

This work introduces a country-level exposure metric derived from a firm-level cascade model (adapted DebtRank) to quantify international systemic risk flows via supply networks. It shows that exposures are regionally structured, asymmetric across income groups, and highly unequal globally, with poorer countries disproportionately exposed and no evident growth compensation. The study highlights the need for granular network monitoring and systemic risk-aware policies.

Future research should incorporate product types and firm-level production functions, detailed link weights and firm sizes (revenues, volumes), empirical heterogeneous default probabilities and correlated shocks, and dynamic features such as inventories, substitution, and recovery. Developing unified measures capturing both upstream and downstream risks and undertaking empirical validation against observed post-shock outcomes would strengthen the framework. Investigating how firms form cross-border ties and designing incentives to reduce exposure burdens on poorer countries are further priorities.

• Product heterogeneity and production functions are not modeled; the cascade uses effectively linear production assumptions, likely underestimating real shock propagation in some cases.
• Incomplete coverage: the constructed network includes N=230,970 firms—far fewer than the global total—raising sampling concerns, though sampling tests show high correlation with full-network results.
• Missing economic weights: no firm revenues or traded volumes; degree is used as a proxy for firm size and link weights are unweighted. Robustness checks (e.g., Ecuador VAT data, removing low-value links) suggest qualitative robustness.
• Potential reporting bias toward U.S.-focused firms in the dataset; analogous analyses within the U.S. show similar qualitative patterns.
• Upstream and downstream shocks are analyzed separately; a unified metric capturing both simultaneously is desirable.
• The cascade disregards inventories, substitution, and recovery; a simple substitution mechanism tested in SI preserves key patterns, but richer dynamics are not modeled.
• Default probabilities are not included in main results; SI indicates robustness to heterogeneous simulated PDs, but empirical PDs and correlated shocks are lacking.
• Limited external validation of cascade outcomes due to data constraints; comprehensive validation against observed revenue/output changes after exogenous shocks remains future work.