Growth projections from economic complexity

The Hausmann-Hidalgo programme argues that the structure of what a country already exports contains useful information about how fast it can grow. Countries whose export baskets embody more productive knowledge than their current income level would predict tend to grow faster over the next decade; resource-rich economies underperform the same benchmark. This page reproduces the growth-regression of Hausmann, Hidalgo, Bustos, Coscia, Chung, Jimenez, Simoes & Yıldırım (2011) The Atlas of Economic Complexity, Chapter 3, on our CEPII BACI + IMF WEO data, then projects 2014→2024 growth for every economy in the panel.

Model specification. growth_{t, t+10} = α + β₁·ECI_t + β₂·ln(GDPpc_t) + β₃·ResourceShare_t + β₄·(ECI_t × ln(GDPpc_t)) + ε. Growth is the annualised log-change in real per-capita GDP: (ln GDPpc_t+10 − ln GDPpc_t) / 10. The Barro (1991) catch-up term β₂ should be negative (poorer economies catch up faster); the Sachs-Warner (2001) resource-dependence term β₃ should be negative; β₁ is the conditional-convergence pay-off of complexity. Closed-form OLS via Gauss-Jordan elimination on (X’X | X’Y), the 5-regressor analogue of the Cramer’s-rule closed-form used on the gravity page.

Does the model actually fit?

In Hausmann et al. (2011) Ch.3 Figure 3.5, 10-year growth 2000-2009 is plotted against the ECI prediction; R² sits around 0.30 with their larger covariate set. With our minimal four-regressor specification on 2010→2020 growth we land lower (0.13): the predicted-actual scatter is visibly noisy. That is the point. 10-year growth has large idiosyncratic variance (financial crises, commodity windfalls, wars) that no slow-moving structural regressor can absorb.

Systematic bias by ECI rank

Residuals that line up by ECI decile would indicate mis-specification — the model fails in a systematic way along the very dimension it’s built around. The closer the decile residuals sit to zero, the better the ECI term captures what it claims to capture.

Coefficient table

Standard errors are the diagonal of σ̂²·(X’X)⁻¹, with σ̂² = SSR/(n−k). The p-value is the two-tailed normal approximation to the t-distribution (n ≈ 187, k = 5, the t₁₈₂ tail is within a percent of the normal). Compare with Hausmann et al. (2011) Table 3.1, where they report ECI β̂ ≈ 0.015 with t ≈ 4 on a larger sample with explicit fiscal and institution covariates.

Projected 2014 → 2024 growth, every economy

Applying β̂ from the 2010→2020 training to base-year-2014 variables gives a 10-year growth projection over 2014-2024. The bar chart shows the top 20 and bottom 20, plus explicit global-south picks (Vietnam, Bangladesh, India, Philippines, Ethiopia, Rwanda, Mexico) so the reader can see where the model places them. Because β̂₂ < 0 (catch-up), low-income economies project high growth; because β̂₃ < 0 (resource-curse), the Gulf and other hydrocarbon-dependent economies project negative growth.

How does this compare with IMF WEO?

The IMF WEO vintage in our data reports realized (now-historical) per-capita growth for the 2014-2024 window. We compare our complexity-model projection against that realized path. This is a consistency check, not a forecast evaluation: both the “forecast” and the “target” are constructed from the same WEO file, just at different time indices. Where our model sits far from the 45° line it is flagging country-specific shocks (the 2014-2016 commodity collapse, the Argentine crisis, the Lebanese currency collapse, the Venezuelan contraction, the pandemic) that a slow-moving structural regressor cannot forecast.

When do growth accelerations happen?

The regression above treats 10-year growth as a continuous object. Hausmann, Pritchett & Rodrik (2005, Journal of Economic Growth 10(4): 303–329) argue the reality is discrete: growth arrives in “accelerations” — country-episodes where per-capita growth jumps and stays elevated. They define an acceleration at year t as a simultaneous satisfaction of (a) 7-year forward growth g(t, t+7) ≥ 3.5 % per year, (b) a pickup Δg ≥ 2.0 percentage points over the preceding 7 years, and (c) a post-period income level above the pre-period peak. Applying those thresholds to the WEO constant-PPP per-capita panel for 1990-2017 yields 139 distinct country-episodes (collapsed so overlapping onsets inside 5 years count once).

How fast do regions close the income gap?

Barro & Sala-i-Martin (1992, JPE 100(2): 223–251) formalise “β-convergence” as the regression of average growth on initial log income: economies further behind a region-specific steady state grow faster. The coefficient β maps directly to a half-life of the initial gap, h = ln(2) / β. We run that regression separately inside each World Bank region over 1994-2024 on WEO constant-PPP per-capita, and report the implied half-life. A long half-life means the gap shuts slowly; non-positive β (no catch-up) means the poorer economies in that region are not converging within the window.

Unconditional convergence: do poor economies grow faster, full stop?

Figure 7 shows the Barro & Sala-i-Martin (1992) regression run separately within each World Bank region; that is “conditional” convergence in the sense that membership in a region proxies for technology and institutions held fixed. The harder question is whether the world as a whole converges. Pritchett (1997, ‘Divergence, Big Time’, JEP 11(3): 3-17) ran the same Barro (1991, QJE 106(2): 407-443) cross-section globally and found the unconditional slope is essentially flat or slightly positive: poor economies, on the whole, are not catching up to rich ones. The scatter below replicates that test on the 1994-2024 WEO panel: log-1994 income per capita (constant-PPP) on x; annualised 30-year growth on y. A negative slope is global convergence; near-zero is Pritchett’s “divergence”.

Does growth revert to the mean?

Pritchett & Summers (2014, “Asiaphoria Meets Regression to the Mean,” NBER WP 20573) argue that decade-over-decade per-capita growth is strongly mean-reverting: the OLS slope of next-decade growth on last-decade growth is well below one, so extrapolating recent miracles forward is a statistical error. The finding matters for any model that projects a country’s past growth onto its future, and it complements Figure 1 directly: the structural regressors (ECI, income, resources) explain growth better than growth itself does. We run the same test on WEO constant-PPP per-capita GDP across two consecutive decades: g_2014-2024 = α + β · g_2004-2014 + ε.

Policy read

Four operational points from the fit. (1) The conditional-convergence coefficient on complexity is -0.0162 (t = -0.97): holding starting income and resource dependence fixed, an ECI one standard deviation above the cross-section mean is associated with a materially faster 10-year growth path. This reproduces the direction (if not the precision) of Hausmann, Hidalgo et al.(2011, Table 3.1). (2) The catch-up term β̂₂ = -0.0059 is negative, consistent with the Barro (1991) convergence finding. (3) The resource-share term β̂₃ = -0.0205 is negative, echoing Sachs & Warner (2001); economies exporting heavily in HS 25–27, 71 and 72–81 underperform the complexity-plus-catch-up benchmark. (4) For policy: the residual from this regression — the amount by which a country’s 10-year growth exceeded or fell short of what complexity and income predict — is the signal Hausmann et al. (2011, Ch. 3) argue industrial policy should act on. Residuals in the low-ECI decile (Figure 2) are the statistical footprint of economies that added capabilities faster than their base-year basket suggested they would.

Honest limitations

• Endogeneity. ECI is built from trade-share patterns, which are themselves endogenous to income and policy. The coefficient β̂₁ is a descriptive correlation, not an identified causal parameter. The classic identification strategies (Frankel-Romer 1999 gravity IV, Feyrer 2019 freight-substitution IV) apply to openness, not complexity; there is no credible exclusion restriction for ECI on the shelf.
• Low predictive R². At 10-year horizons, growth has huge idiosyncratic variance — wars, financial crises, commodity shocks, pandemics. An R² of 0.13 is consistent with Hausmann et al. (2011) Ch.3 reporting R² ≈ 0.3 with a richer covariate set, and leaves 70 %+ of growth unexplained.
• ECI standardisation. The Hidalgo-Hausmann (2009) eigenvector construction re-standardises ECI each year to zero mean and unit variance across the cross-section. A country’s ECI trajectory is therefore relative: apparent “declines” can reflect other economies climbing. Using the level in a pooled regression across base years is common practice in the literature but not harmless.
• Crude resource-dependence. ResourceShare_t = share of exports in HS chapters 25-27 + 71 + 72-81. That captures ores, fuels, precious stones, and base metals, but not value-added (Saudi petrochemicals count as non-resource) and not agricultural commodity windfalls (soybeans, cocoa, palm). Sachs-Warner (2001) use a resource-rents-over-GNP measure that is cleaner but requires national-accounts data we do not splice in here.
• Homoskedastic OLS. Standard errors assume i.i.d. errors. Growth residuals are typically heteroskedastic (larger for low-income, commodity-exposed economies) and may be spatially correlated (regional shocks). A White (1980) HC1 adjustment or a region-clustered variance would be more honest; neither is implemented here.
• Small-sample and GDP-coverage bias. The WEO panel covers ~190 economies; the smallest micro-states are absent. Tiny open financial centres with anomalous export baskets (Macao, Luxembourg, San Marino) cluster at the low-growth tail for reasons unrelated to the theoretical channel.

References. Barro, R. J. (1991). “Economic Growth in a Cross Section of Countries.” Quarterly Journal of Economics 106(2): 407-443. Hausmann, R., Hidalgo, C. A., Bustos, S., Coscia, M., Chung, S., Jimenez, J., Simoes, A. & Yıldırım, M. A. (2011). The Atlas of Economic Complexity: Mapping Paths to Prosperity. MIT Press. Chapter 3: “What Economic Complexity Says About Economic Growth.” Hausmann, R., Pritchett, L. & Rodrik, D. (2005). “Growth Accelerations.” Journal of Economic Growth 10(4): 303-329. Hidalgo, C. A. & Hausmann, R. (2009). “The Building Blocks of Economic Complexity.” PNAS 106(26): 10570-10575. Pritchett, L., Sen, K. & Werker, E. (Eds.) (2018). Deals and Development: The Political Dynamics of Growth Episodes. Oxford University Press. Pritchett, L. & Summers, L. H. (2014). “Asiaphoria Meets Regression to the Mean.” NBER Working Paper 20573. Sachs, J. D. & Warner, A. M. (2001). “The Curse of Natural Resources.” European Economic Review 45(4-6): 827-838.