## Section 3.3)

## Source details

**Canonical URL:** [Section 3.3)](https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022075-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2022/english/wpiea2022075-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2022/english/wpiea2022075-print-pdf.pdf.json)

---

### Method: similarity score construction and timing
- Portfolio structure of country i in time t: number of high RCA exports (SITC 4-digit) that belong to each SITC 1-digit sector as a share of the total number of high RCA exports.
- Definitions:
  - Nactual_l,it = number of high RCA exports in sector l.
  - sactual_l,it ≡ Nactual_l,it / ∑_{l′} N_{l′,it}.
  - Sactual_it ≡ [sactual_l,it]_{L×1}, where L = 10.
  - Srec_it ≡ [srec_l,it]_{L×1}, defined analogously for the recommended product set.
- Similarity score:
  - Sim_it ≡ (Srec_{i,t−Δt} − \bar Srec_{i,t−Δt}) · (Sactual_it − \bar Sactual_i,t) / (‖Srec_{i,t−Δt} − \bar Srec_{i,t−Δt}‖ ‖Sactual_it − \bar Sactual_i,t‖)
  - Baseline time lag Δt = 5 years; alternative Δt choices explored in robustness checks.
- Implementation detail:
  - Annual Sim_it calculated for all country-year pairs; similarity scores and most controls are annual averages over the Δt window. Regressions run with non-overlapping Δt as the time unit.

### Context: RCA, recommendation algorithm, and K selection
- Revealed Comparative Advantage (RCA):
  - RCA_{ij} = E_{ij}/∑_j E_{ij}  ÷  ∑_i E_{ij}/∑_{i}∑_{j} E_{ij}, where E_{ij} is export value of product j from country i.
  - High‑RCA product defined as RCA_{ij} > 1.
- Number of recommended products K:
  - K_rca,it = β_1 GDP_it + β_2 POP_it + γ_t + ε_it
  - GDP_it is GDP per capita (relative to U.S. level in analysis described); POP_it is population size; γ_t is time fixed effect.
  - ˆK_rca,it is the predicted value used as K.
- Recommendation algorithm:
  - Product-based K-nearest neighbor (KNN) using cosine similarity between product vectors p_j (RCA scores across m country-year observations).
  - For country i, high‑RCA set = {j | r_ij > 1}.
  - Predicted r_ij for any product j = weighted average RCA of the country’s high‑RCA set, weights = cosine similarity between product j and products in the high‑RCA set.
  - Recommended products = top ˆK_rca,it products with highest predicted r_ij.
- Training set:
  - Cross-sectional data for a single year (y = 1); when generating recommendations for 2017, only cross-country export data for 2017 is included.

### Empirical use of similarity scores and estimators
- Main regressions use Sim_it as key explanatory variable for growth, volatility, and risk-adjusted growth.
- Estimators: OLS, country & time fixed effects, and system GMM (Arellano and Bond (1991)).
  - In system GMM: lagged dependent and country-level controls treated as endogenous and instrumented; time fixed effects and world-level controls treated as exogenous.

### Key econometric specifications where Sim_it is included
- Growth regression:
  - g_it = β_0 + β_1 y_{i,t−Δt} + β_2 Sim_it + γ X_it + ε_it
  - g_it = average annual growth in GDP per capita over the Δt period; y_{i,t−Δt} = lagged real GDP per capita (log); X_it includes investment-to-GDP ratio, human capital growth, TFP growth, and world GDP growth in some specifications.
  - Main parameter of interest: β_2.
- Volatility regression:
  - vol_it = β_0 + β_1 vol_{i,t−Δt} + β_2 Sim_it + γ X_it + ε_it
  - vol_it = standard deviation of annual growth of real GDP per capita during the Δt period; world growth volatility included in some specifications.
- Risk-adjusted growth regression:
  - g^ra_it = β_0 + β_1 g^ra_{i,t−Δt} + β_2 Sim_it + γ X_it + ε_it
  - g^ra_it ≡ (g_it − g_t) / σ_it over the Δt period.

### Summary statistics relevant to similarity and outcomes (from Table 1)
- Similarity score: Mean 0.817, Std. Dev. 0.179, Min -0.203, Max 0.995, N 1203
- GDP per capita: Mean 8.390, Std. Dev. 1.505, Min 5.129, Max 11.663, N 1414
- Investment rate: Mean 0.219, Std. Dev. 0.109, Min -0.479, Max 0.942, N 1398
- TFP growth: Mean 0.002, Std. Dev. 0.028, Min -0.184, Max 0.222, N 868
- Human capital growth: Mean 0.01, Std. Dev. 0.007, Min -0.025, Max 0.043, N 1107
- GDP per capita growth: Mean 0.017, Std. Dev. 0.037, Min -0.247, Max 0.367, N 1324
- Growth volatility: Mean 0.015, Std. Dev. 0.014, Min 0.001, Max 0.142, N 1177
- Risk-adjusted growth: Mean 0.705, Std. Dev. 4.474, Min -75.438, Max 50.674, N 1175

### Main empirical findings on Sim_it
- Baseline (Δt = 5 years) produces annual similarity scores used in over 700 observations for GMM estimation.
- Growth:
  - Similarity score is positively and significantly associated with GDP per capita growth across OLS, fixed-effect, and system GMM specifications.
  - System GMM (Column 5, baseline growth regression): a 0.1 increase in Sim_it is associated with a 0.22 percentage point increase in the annual growth rate of real GDP per capita; coefficient significant at 1% level. This is equivalent to a move from the median to the 90th percentile of the similarity score distribution.
  - Reported Growth regression coefficients for Similarity Score: 0.023***, 0.036***, 0.035**, 0.023***, 0.022*** (standard errors (0.008)(0.011)(0.014)(0.006)(0.006)).
- Volatility:
  - Similarity score is negatively and significantly associated with growth volatility in two OLS and two system GMM specifications; negative but not significant in fixed-effect estimation.
  - Example baseline coefficient: a 0.1 increase in Sim_it is associated with a 0.0015 decrease in the standard deviation of growth rate in a 5-year window (statistically significant at 1% level).
  - Volatility regression coefficients for Similarity Score: -0.010***, -0.023**, -0.022, -0.014**, -0.015*** (standard errors (0.004)(0.011)(0.014)(0.006)(0.006)).
- Risk-adjusted growth:
  - System GMM (Column 5): a 0.1 increase in Sim_it is associated with 20.95 percentage points increase per standard deviation in the annual growth rate; significant at 5% level.
  - Reported Risk-adjusted growth coefficients for Similarity Score: 3.406***, 3.270, 2.604, 1.985**, 2.095** (standard errors (0.627)(2.589)(2.568)(0.863)(0.853)).
- Additional numeric illustrations:
  - Baseline: a 0.5 increase in the similarity score is associated with a 1.1 percentage point increase in the annual growth of real GDP per capita, and a 0.0075 decrease in the standard deviation of growth rate over 5-year time windows.
  - The risk-adjusted growth variable defined as average annual growth divided by the standard deviation of growth during Δt period.
  - System GMM: an increase of 0.8 in the similarity score moves a country from the world median in the risk-adjusted growth spectrum to the 75th percentile level.

### Robustness and sensitivity results
- Time-interval sensitivity (system GMM with controls):
  - Growth: ∆t = 3 → 0.1 increase in Sim_it associated with 0.19 percentage point increase in growth.
  - Growth: ∆t = 7 → 0.1 increase associated with 0.13 percentage point increase.
  - Volatility: ∆t = 3 → 0.1 increase associated with 0.0017 decrease in standard deviation; ∆t = 7 → 0.0013 decrease.
  - Risk-adjusted growth: ∆t = 3 → 0.1 increase associated with 21.84 percentage point increase (1% significance); ∆t = 7 → 13.79 (not statistically significant).
- Winsorization and outliers:
  - Winsorized growth regression (top and bottom 5% removed): similarity coefficient remains positive and significant; system GMM indicates a 0.1 increase associated with 0.18 percentage point increase in real GDP per capita growth.
  - Winsorized volatility regression: similarity coefficients remain negative but weaker; not significant in four out of five specifications.
  - Winsorized risk-adjusted growth regression: system GMM shows a 0.1 increase associated with 9.8 percentage points increase per standard deviation; significance shifts from 1% to 10%.
- Forward-looking regressions (selected coefficients):
  - Growth (fwd+3): Similarity Score estimates include 0.025***, 0.022**, 0.023**, 0.020***, 0.019*** (standard errors (0.007)(0.010)(0.010)(0.007)(0.007)).
  - Growth (fwd+7): Similarity Score estimates include 0.015**, 0.019, 0.021**, 0.014**, 0.013** (standard errors (0.007)(0.015)(0.010)(0.007)(0.007)).
  - Volatility (fwd+3): Similarity Score estimates include -0.014***, -0.011, -0.011, -0.015**, -0.017*** (standard errors (0.003)(0.009)(0.011)(0.006)(0.005)).
  - Risk-adjusted Growth (fwd+3): Similarity Score estimates include 2.641***, 2.259*, 2.870**, 2.403***, 2.184*** (standard errors (0.418)(1.167)(1.311)(0.824)(0.679)).

### Regression diagnostics and significance conventions
- Significance notation: * p <0.1, ** p <0.05, *** p <0.01.
- Diagnostic tests frequently reported: AR1 (p-value), AR2 (p-value), Hansen-J (p-value).
- Selected diagnostics from baseline tables:
  - Growth (Table 2): AR1 (p-value): 0.00, 0.00; AR2 (p-value): 0.01, 0.02; Hansen-J (p-value): 0.79, 0.80.
  - Volatility (Table 3): AR1 (p-value): 0.00, 0.00; AR2 (p-value): 0.51, 0.53; Hansen-J (p-value): 0.58, 0.38.
  - Risk-adjusted growth (Table 4): AR1 (p-value): 0.01, 0.02; AR2 (p-value): 0.49, 0.17; Hansen-J (p-value): 0.49, 0.43.

### Robustness: descriptive distributions and classifications
- Actual RCA by year (selected):
  - 1985: N 611600, Mean 0.136, Q1 0.024, Median 0.158, Q3 0.844
  - 2015: N 1017890, Mean 0.117, Q1 0.022, Median 0.143, Q3 0.708
- Recommendation scores by year (selected):
  - 1985: N 1337220, Mean 0.269, Q1 0.140, Median 0.328, Q3 0.648
  - 2015: N 1605760, Mean 0.337, Q1 0.194, Median 0.393, Q3 0.702
- SITC 1-section mapping (Table 17):
  - 0: Food and live animals
  - 1: Beverages and tobacco
  - 2: Crude materials, inedible, except fuels
  - 3: Mineral fuels, lubricants and related materials
  - 4: Animal and vegetable oils, fats and waxes
  - 5: Chemicals and related products, n.e.s.
  - 6: Manufactured goods classified chiefly by material
  - 7: Machinery and transport equipment
  - 8: Miscellaneous manufactured articles
  - 9: Commodities and transactions not classified elsewhere in the SITC

### Interpretation and policy-relevant conclusions
- Methodological contribution:
  - Combines machine learning (product-based KNN at SITC 4-digit level for over 190 economies, 1980–2018) to generate export-product recommendations with standard linear regressions to evaluate whether greater alignment between recommended and actual export structures correlates with better growth outcomes.
- Practical implications:
  - Algorithm-produced export recommendations appear to be a useful diagnostic tool for policymakers and private investors evaluating industrial policy options and market entry, but they are not substitutes for detailed, multidimensional analyses of industry viability.
  - Knowing potential comparative advantages does not automatically translate into specific policy interventions; policy design must be country-specific and consider institutional constraints.
  - The paper can inform broader reform agendas and product diversification strategies for small open economies and large emerging markets, serving as a diagnostic input for prioritizing reforms and export diversification.
- Caveats:
  - Authors do not advocate direct policy interventions solely based on the recommendations; results are diagnostic and indicative of directions for further investigation.

*Source: wpiea2022075-print-pdf - Section 3.3).*

### Section 3.3).

### Section 3.3)

### Method: similarity score construction and timing
- Portfolio structure of country i in time t is defined as the number of high RCA exports (SITC 4-digit) that belong to each SITC 1-digit sector as a share of the total number of high RCA exports.
- Let Nactual_l,it be the number of high RCA exports in sector l; sactual_l,it ≡ Nactual_l,it / ∑_{l′} Nactual_{l′,it}. Actual export structure Sactual_it ≡ [sactual_l,it]_{L×1}, where L = 10.
- Recommended export structure Srec_it ≡ [srec_l,it]_{L×1}, defined analogously for the recommended product set.
- Similarity score Sim_it is computed as the cosine similarity between demeaned recommended and actual sector-share vectors:
  - Sim_it ≡ (Srec_{i,t−Δt} − \bar Srec_{i,t−Δt}) · (Sactual_it − \bar Sactual_i,t) / (‖Srec_{i,t−Δt} − \bar Srec_{i,t−Δt}‖ ‖Sactual_it − \bar Sactual_i,t‖)
- Time lag Δt is used because export structure evolves slowly and recommendations are somewhat forward-looking; baseline estimation sets Δt = 5 years. Alternative Δt choices are explored in robustness checks.

### Context: RCA, recommendation algorithm, and K selection
- Revealed Comparative Advantage (RCA) for country i in product j: RCA_{ij} = E_{ij}/∑_j E_{ij}  ÷  ∑_i E_{ij}/∑_{i}∑_{j} E_{ij}, where E_{ij} is export value of product j from country i.
- A high‑RCA product for country i is defined as RCA_{ij} > 1.
- Number of recommended products K is set to the predicted number of high‑RCA exports ˆK_rca,it from the regression:
  - K_rca,it = β_1 GDP_it + β_2 POP_it + γ_t + ε_it
  - GDP_it is GDP per capita (relative to U.S. level in analysis described); POP_it is population size; time fixed effect γ_t included. ˆK_rca,it is the predicted value used as K.
- Recommendation algorithm:
  - Product-based K-nearest neighbor (KNN) using cosine similarity between product vectors p_j (RCA scores across m country-year observations).
  - For country i, high‑RCA set = {j | r_ij > 1}.
  - Predicted r_ij for any product j is the weighted average RCA of the country’s high‑RCA set, weights = cosine similarity between product j and products in the high‑RCA set.
  - Recommended products = top ˆK_rca,it products with highest predicted r_ij.
- Training set in practice uses cross-sectional data for a single year (y = 1); when generating recommendations for 2017, only cross-country export data for 2017 is included.

### Empirical use of similarity scores
- Annual Sim_it calculated for all country-year pairs and incorporated into subsequent regressions on growth, volatility, and risk-adjusted growth.
- Baseline Δt = 5 years; similarity scores and most controls are annual averages over the Δt window. Regressions run with non-overlapping Δt as the time unit.

### Key econometric specifications where Sim_it is included
- Growth regression:
  - g_it = β_0 + β_1 y_{i,t−Δt} + β_2 Sim_it + γ X_it + ε_it
  - g_it = average annual growth in GDP per capita over the Δt period; y_{i,t−Δt} = lagged real GDP per capita (log); X_it includes investment-to-GDP ratio, human capital growth, TFP growth, and world GDP growth in some specifications; country and time fixed effects included in some specs. Main parameter of interest: β_2.
- Volatility regression:
  - vol_it = β_0 + β_1 vol_{i,t−Δt} + β_2 Sim_it + γ X_it + ε_it
  - vol_it = standard deviation of annual growth of real GDP per capita during the Δt period; world growth volatility included in controls for some specifications.
- Risk-adjusted growth regression:
  - g^ra_it = β_0 + β_1 g^ra_{i,t−Δt} + β_2 Sim_it + γ X_it + ε_it
  - g^ra_it ≡ (g_it − g_t) / σ_it over the Δt period; X_it replaces world growth with average risk-adjusted growth across countries in some specifications.
- Estimators used: OLS, country & time fixed effects, and system GMM (Arellano and Bond (1991)) to address endogeneity from lagged dependent variables. In system GMM, lagged dependent and country-level controls treated as endogenous and instrumented; time fixed effects and world-level controls treated as exogenous.

### Summary statistics relevant to similarity and outcomes (from Table 1)
- Similarity score: Mean 0.817, Std. Dev. 0.179, Min -0.203, Max 0.995, N 1203
- GDP per capita: Mean 8.390, Std. Dev. 1.505, Min 5.129, Max 11.663, N 1414
- Investment rate: Mean 0.219, Std. Dev. 0.109, Min -0.479, Max 0.942, N 1398
- TFP growth: Mean 0.002, Std. Dev. 0.028, Min -0.184, Max 0.222, N 868
- Human capital growth: Mean 0.01, Std. Dev. 0.007, Min -0.025, Max 0.043, N 1107
- GDP per capita growth: Mean 0.017, Std. Dev. 0.037, Min -0.247, Max 0.367, N 1324
- Growth volatility: Mean 0.015, Std. Dev. 0.014, Min 0.001, Max 0.142, N 1177
- Risk-adjusted growth: Mean 0.705, Std. Dev. 4.474, Min -75.438, Max 50.674, N 1175

### Empirical findings on Sim_it (high-level)
- Baseline Δt = 5 years yields annual similarity scores used in over 700 observations for GMM estimation.
- Similarity score is positively and significantly associated with GDP per capita growth across OLS, fixed-effect, and system GMM specifications.
  - System GMM (Column 5, baseline growth regression): a 0.1 increase in Sim_it is associated with a 0.22 percentage point increase in the annual growth rate of real GDP per capita; coefficient significant at 1% level. This is equivalent to a move from the median to the 90th percentile of the similarity score distribution.
- Similarity score is negatively and significantly associated with growth volatility in two OLS and two system GMM specifications (i.e., higher alignment reduces volatility); coefficient is negative but not significant in the fixed-effect estimation.
- Control variables:
  - Investment rate, TFP growth, and human capital growth are generally positive and mostly significant in growth regressions.
  - Lagged GDP per capita is negative and significant in OLS and fixed-effect estimations (consistent with convergence), but negative and insignificant in system GMM.
  - Lagged volatility is positive and significant across specifications, indicating persistence in growth volatility.
  - World growth volatility is positive and significant in volatility regressions, indicating external environment matters for country volatility.

*Source: wpiea2022075-print-pdf - Section 3.3).*

### 0.1 increase in the similarity score is associated with a 0.0015 decrease in the standard deviation

### wpiea2022075-print-pdf - 0.1 increase in the similarity score is associated with a 0.0015 decrease in the standard deviation

### Main empirical findings
- Growth and volatility:
  - Baseline: a 0.1 increase in the similarity score is associated with a 0.0015 decrease in the standard deviation of growth rate in a 5-year window. This is statistically significant at 1% level.
  - Baseline: a 0.5 increase in the similarity score is associated with a 1.1 percentage point increase in the annual growth of real GDP per capita, and a 0.0075 decrease in the standard deviation of growth rate over 5-year time windows.
- Risk-adjusted growth:
  - The risk-adjusted growth variable is defined as the average annual growth divided by the standard deviation of growth during ∆t period.
  - System GMM (Column 5): a 0.1 increase in the similarity score is associated with 20.95 percentage points increase per standard deviation in the annual growth rate. Result is statistically significant at 5% level.
  - An increase of 0.8 in the similarity score moves a country from the world median in the risk-adjusted growth spectrum to the 75th percentile level.
- Controls and dynamics:
  - The lagged dependent variable has a positive coefficient in all specifications, significant in the OLS without controls and the fixed effect specification.
  - Among controls, investment rate and TFP growth are associated with higher risk-adjusted growth.
  - The human capital variable is insignificant and its sign varies across estimations.
  - The dependent variable is strongly correlated with world-level risk-adjusted growth.

### Robustness: time-interval sensitivity (∆t)
- Growth regressions (system GMM with controls):
  - ∆t = 3: a 0.1 increase in the similarity score is associated with 0.19 percentage point increase in growth.
  - ∆t = 7: a 0.1 increase in the similarity score is associated with 0.13 percentage point increase in growth.
- Volatility regressions (system GMM with controls):
  - ∆t = 3: a 0.1 increase in the similarity score is associated with 0.0017 decrease in the standard deviation of growth.
  - ∆t = 7: a 0.1 increase in the similarity score is associated with 0.0013 decrease in the standard deviation of growth.
- Risk-adjusted growth (system GMM with controls):
  - ∆t = 3: a 0.1 increase in the similarity score is associated with 21.84 percentage point increase in the risk-adjusted annual growth rate, at 1% significance level.
  - ∆t = 7: the corresponding value is 13.79 and it is not statistically significant.
- Summary visualization:
  - Figure 9 (referenced) summarizes estimates with time intervals varying between 3 to 7 and shows confidence intervals for key coefficients in growth, volatility, and risk-adjusted growth regressions.

### Robustness: winsorization and outliers
- Winsorized growth regression (top and bottom 5% removed):
  - The similarity score coefficient remains positive and significant across all specifications.
  - System GMM (Column 5) indicates a 0.1 increase in the similarity score is associated with a 0.18 percentage point increase in real GDP per capita growth.
- Winsorized volatility regression:
  - Similarity score coefficients remain negative in all five specifications but are weaker in magnitude and not significant in four out of five specifications.
- Winsorized risk-adjusted growth regression:
  - System GMM shows a 0.1 increase in the similarity score is associated with 9.8 percentage points increase per standard deviation in the annual growth rate.
  - This is smaller than the baseline 20.95 percentage points and statistical significance shifts from 1% to 10%.
- Overall:
  - Winsorized results still confirm a positive correlation between the similarity score and comprehensive growth performance, though magnitudes and significance levels are reduced.

### Interpretation and policy-relevant conclusions
- Methodological contribution:
  - The paper combines machine learning (product-based KNN at SITC 4-digit level for over 190 economies, 1980–2018) to generate export-product recommendations with standard linear regressions to evaluate whether greater alignment between recommended and actual export structures correlates with better growth outcomes.
- Practical implications:
  - Algorithm-produced export recommendations appear to be a useful diagnostic tool for policymakers and private investors evaluating industrial policy options and market entry, but they are not substitutes for detailed, multidimensional analyses of industry viability.
  - Knowing potential comparative advantages does not automatically translate into specific policy interventions; policy design must be country-specific and consider institutional constraints.
  - The paper can inform broader reform agendas and product diversification strategies for small open economies and large emerging markets, serving as a diagnostic input for prioritizing reforms and export diversification.
- Caveats:
  - The authors do not advocate direct policy interventions solely based on the recommendations; results are diagnostic and indicative of directions for further investigation.

*Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022075-print-pdf.pdf*

### References

### References

### Regression Results — Key Coefficients (baseline and main specifications)
- Growth regression (Table 2):
  - Lagged GDP Per Capita: -0.002**, -0.008***, -0.016***, -0.001, -0.001 (standard errors (0.001)(0.003)(0.004)(0.001)(0.001))
  - Similarity Score: 0.023***, 0.036***, 0.035**, 0.023***, 0.022*** (standard errors (0.008)(0.011)(0.014)(0.006)(0.006))
  - Inv. Rate: 0.016***, 0.014***, 0.021***, 0.021*** (standard errors (0.002)(0.002)(0.004)(0.004))
  - TFP Growth: 0.716***, 0.689***, 0.745***, 0.742*** (standard errors (0.063)(0.057)(0.064)(0.061))
  - Human Capital Growth: 0.426***, 0.420***, 0.600***, 0.571*** (standard errors (0.127)(0.133)(0.149)(0.153))
  - World Growth: 0.381*, 0.263 (standard errors (0.195)(0.193))
  - No. of Obs.: 1168, 732, 273, 273, 2732
  - AR1 (p-value): 0.00, 0.00; AR2 (p-value): 0.01, 0.02; Hansen-J (p-value): 0.79, 0.80

- Volatility regression (Table 3):
  - Lagged Volatility: 0.443***, 0.272***, 0.268***, 0.468***, 0.468*** (standard errors (0.071)(0.075)(0.059)(0.056)(0.055))
  - Similarity Score: -0.010***, -0.023**, -0.022, -0.014**, -0.015*** (standard errors (0.004)(0.011)(0.014)(0.006)(0.006))
  - Inv. Rate: -0.001, -0.001, -0.003**, -0.002 (standard errors (0.001)(0.001)(0.001)(0.001))
  - TFP Growth: -0.011, -0.013, -0.012, -0.011 (standard errors (0.035)(0.026)(0.023)(0.027))
  - Human Capital Growth: 0.136*, 0.138, 0.112, 0.082 (standard errors (0.080)(0.090)(0.083)(0.078))
  - World Volatility: 0.675***, 0.464** (standard errors (0.190)(0.209))
  - No. of Obs.: 9666, 1461, 4613, 614, 614
  - AR1 (p-value): 0.00, 0.00; AR2 (p-value): 0.51, 0.53; Hansen-J (p-value): 0.58, 0.38

- Risk-adjusted growth regression (Table 4):
  - Lagged Adj. Growth: 0.311***, 0.073, 0.086*, 0.066, 0.050 (standard errors (0.111)(0.045)(0.048)(0.052)(0.046))
  - Similarity Score: 3.406***, 3.270, 2.604, 1.985**, 2.095** (standard errors (0.627)(2.589)(2.568)(0.863)(0.853))
  - Inv. Rate: 1.989***, 2.247***, 1.775***, 1.721*** (standard errors (0.487)(0.527)(0.482)(0.541))
  - TFP Growth: 33.794***, 38.395***, 36.424***, 29.998*** (standard errors (8.458)(11.167)(10.365)(8.329))
  - Human Capital Growth: 1.194, -11.629, -25.097, -14.007 (standard errors (29.750)(28.030)(34.874)(34.170))
  - World Adj. Growth: 0.508**, 0.578*** (standard errors (0.218)(0.215))
  - No. of Obs.: 9646, 1361, 3612, 613, 613
  - AR1 (p-value): 0.01, 0.02; AR2 (p-value): 0.49, 0.17; Hansen-J (p-value): 0.49, 0.43

### Robustness and Time-horizon Results (selected)
- Robustness: Growth (Table 5, horizons +3, +5, +7):
  - Similarity Score: 0.019***, 0.022***, 0.013** (standard errors (0.007)(0.006)(0.007))
  - Inv. Rate: 0.021***, 0.021***, 0.018*** (standard errors (0.003)(0.004)(0.004))
  - TFP Growth: 0.797***, 0.742***, 0.758*** (standard errors (0.067)(0.061)(0.076))
  - Human Capital Growth: 0.650***, 0.571***, 0.530*** (standard errors (0.158)(0.153)(0.183))
  - World Growth: 0.391***, 0.263, 0.613*** (standard errors (0.085)(0.193)(0.210))
  - No. of Obs.: 1258, 732, 523

- Robustness: Volatility (Table 6, horizons +3, +5, +7):
  - Lagged Volatility: 0.417***, 0.468***, 0.243*** (standard errors (0.058)(0.055)(0.073))
  - Similarity Score: -0.017***, -0.015***, -0.013*** (standard errors (0.005)(0.006)(0.004))
  - TFP Growth: -0.116***, -0.011, 0.088** (standard errors (0.032)(0.027)(0.035))
  - World Volatility: 0.526***, 0.464**, 0.532** (standard errors (0.132)(0.209)(0.234))
  - No. of Obs.: 1231, 614, 409

- Forward-looking regressions (fwd+3 and fwd+7) — selected coefficients:
  - Growth (fwd+3, Table 8): Similarity Score: 0.025***, 0.022**, 0.023**, 0.020***, 0.019*** (standard errors (0.007)(0.010)(0.010)(0.007)(0.007)); No. of Obs.: 2030, 1258, 1258, 1258, 1258
  - Growth (fwd+7, Table 9): Similarity Score: 0.015**, 0.019, 0.021**, 0.014**, 0.013** (standard errors (0.007)(0.015)(0.010)(0.007)(0.007)); No. of Obs.: 838, 523, 522, 523, 523
  - Volatility (fwd+3, Table 10): Similarity Score: -0.014***, -0.011, -0.011, -0.015**, -0.017*** (standard errors (0.003)(0.009)(0.011)(0.006)(0.005)); No. of Obs.: 1948, 1231, 1231, 1231, 1231
  - Risk-adjusted Growth (fwd+3, Table 12): Similarity Score: 2.641***, 2.259*, 2.870**, 2.403***, 2.184*** (standard errors (0.418)(1.167)(1.311)(0.824)(0.679)); No. of Obs.: 1946, 1230, 1230, 1230, 1230

- Winsorized samples (Tables 14–16) maintain:
  - Positive and significant Similarity Score in growth regressions: 0.023***, 0.019*, 0.018*, 0.018***, 0.018*** (Table 14)
  - Negative and often significant Similarity Score in volatility regressions: -0.004*, -0.005, -0.005, -0.008, -0.007 (Table 15)
  - Risk-adjusted growth Similarity Score positive in OLS: 1.878*** (Table 16)

### Descriptive Statistics, Classification, and Distributions
- Summary of Actual RCA by year (Table 18):
  - 1985: N 611600, Mean 0.136, Q1 0.024, Median 0.158, Q3 0.844
  - 1990: N 691300, Mean 0.133, Q1 0.024, Median 0.156, Q3 0.788
  - 1995: N 842130, Mean 0.144, Q1 0.027, Median 0.168, Q3 0.798
  - 2000: N 969250, Mean 0.138, Q1 0.027, Median 0.163, Q3 0.780
  - 2005: N 1009310, Mean 0.128, Q1 0.024, Median 0.155, Q3 0.744
  - 2010: N 1040690, Mean 0.120, Q1 0.023, Median 0.152, Q3 0.735
  - 2015: N 1017890, Mean 0.117, Q1 0.022, Median 0.143, Q3 0.708

- Summary of Recommendation Scores by year (Table 19):
  - 1985: N 1337220, Mean 0.269, Q1 0.140, Median 0.328, Q3 0.648
  - 1990: N 1345040, Mean 0.308, Q1 0.161, Median 0.370, Q3 0.686
  - 1995: N 1522810, Mean 0.350, Q1 0.202, Median 0.403, Q3 0.712
  - 2000: N 1594440, Mean 0.365, Q1 0.217, Median 0.438, Q3 0.754
  - 2005: N 1602180, Mean 0.348, Q1 0.198, Median 0.419, Q3 0.749
  - 2010: N 1607840, Mean 0.342, Q1 0.195, Median 0.416, Q3 0.747
  - 2015: N 1605760, Mean 0.337, Q1 0.194, Median 0.393, Q3 0.702

- SITC 1-section classification (Table 17) — SITC Code to Sector Name mapping:
  - 0: Food and live animals
  - 1: Beverages and tobacco
  - 2: Crude materials, inedible, except fuels
  - 3: Mineral fuels, lubricants and related materials
  - 4: Animal and vegetable oils, fats and waxes
  - 5: Chemicals and related products, n.e.s.
  - 6: Manufactured goods classified chiefly by material
  - 7: Machinery and transport equipment
  - 8: Miscellaneous manufactured articles
  - 9: Commodities and transactions not classified elsewhere in the SITC

### Statistical Significance Conventions and Diagnostic Tests
- Significance notation used throughout tables:
  - * p <0.1, ** p <0.05, *** p <0.01
- Frequently reported diagnostics include:
  - AR1 (p-value), AR2 (p-value), Hansen-J (p-value)
  - Standard errors reported in parentheses for all coefficient estimates

*High Performance Export Portfolio: Design Growth-Enhancing Export Structure with Machine Learning — Working Paper No. WP/2022/075*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022075-print-pdf.pdf_