## sipea2024023

## Source details

**Canonical URL:** [sipea2024023](https://www.imf.org/-/media/files/publications/selected-issues-papers/2024/english/sipea2024023.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/selected-issues-papers/2024/english/sipea2024023.pdf.md)
- [Structured JSON version](/-/media/files/publications/selected-issues-papers/2024/english/sipea2024023.pdf.json)

---

### A. Introduction and Main Messages
- Bulgaria is a small open economy whose integration in global value chains (GVCs) grew before the global financial crisis (GFC) and EU accession, but has since grown only modestly and remains below peers in links with EU countries.
- To harness opportunities from joining the euro and Schengen areas, Bulgaria should improve non-cost competitiveness by improving governance and investing in infrastructure and human capital.
- Authors: Prepared by Giacomo Magistretti and Iglika Vassileva (both EUR).

### B. Developments in Foreign Trade and FDI
- Trade dynamics
  - Bulgaria’s share in world exports steadily increased over the last 3 decades.
  - Exports to the EU rose from 15 percent of GDP in 1999 to 25 percent in 2007, then grew more gradually in the last decade.
  - Notable recent increases in exports to Germany and Romania; declines in importance of Italy, Greece, and Türkiye as export destinations.
  - Imports from the EU have decreased lately, while China, Türkiye, Serbia, and the US increased their importance as suppliers.
  - Exports are shifting toward more technologically-intensive products, with rises in electrical and other machinery and chemicals and declines in textiles and apparel; however, the Economic Complexity Index of exports still lags peers and has not increased over the last two decades.
- Geopolitical energy shifts
  - Import of Russian gas significantly declined after Russia stopped pipeline gas supply to Bulgaria in April 2022.
  - Import of Russian crude oil increased under an EU derogation; a more notable decline in reliance on Russian oil was expected only starting from 2024 after the Bulgarian parliament revoked the derogation from March 1, 2024.
- Foreign Direct Investment (FDI)
  - FDI remains substantially below its 2007 peak and settled to levels similar to other newer member states (NMSs) in the last decade.
  - Recent FDI largely came from EU partners and was mostly concentrated in services, especially financial activities (including real estate), and wholesale and retail trade.
  - Making Bulgaria more attractive to foreign investors is expected to boost integration and attract multinational firms relocating in Europe.

### C. Bulgaria’s Participation in Global Value Chains (GVCs)
- Overall participation and direction
  - In 2020, 35 percent of Bulgaria’s final demand was met by value added coming from abroad, about half of which was from other EU countries.
  - Bulgaria sends a sizeable share of its value added abroad: value added going abroad peaked at almost 45 percent in 2017 and declined to about 38 percent in 2020, close to the EU average.
  - Almost half of value added sent abroad went to other EU countries; the share going to China grew steadily and China became the largest destination outside the EU in 2020, overtaking the US.
- Sectoral composition of value added sent abroad (2020)
  - 25 percent from manufacturing (including construction and utilities).
  - 13 percent from agriculture and mining.
  - 56 percent from services.
  - Shares of wholesale and retail trade and IT and communication in value added sent abroad have been growing steadily over time.
- Labor market interconnectedness
  - Fraction of Bulgarian employment sustained by foreign demand rose from 35 percent during the GFC to about 50 percent by the end of the 2010s, then declined slightly to 46 percent in 2020.
  - Bulgaria’s employment sustained by foreign demand is the fourth highest in the EU behind Ireland, Luxembourg, and Malta.
- Aggregate GVC metrics and trends
  - Bulgaria’s participation in GVCs increased strongly before EU accession and the GFC, declined in 2009, partially recovered, and then increased only marginally thereafter.
  - Bulgaria’s overall GVC participation trails most NMSs and the EU average.
  - Backward linkages (foreign value added embodied in domestic exports) are in line with peers and the EU average.
  - Forward linkages (domestic value added embodied in foreign exports) are limited, below the EU average and below peer countries, though slightly increasing in recent years.
- Industry-level linkages
  - Backward linkages: the largest involvement is in basic metals and coke and refined petrol products, accounting for almost a third of total foreign value added in Bulgaria’s exports in 2020.
  - Forward linkages: well-balanced across manufacturing and services industries.
- Sophistication
  - Involvement in information GVCs (computers, electronics, communication, IT, and information services) is among the lowest in the EU, though sophistication has been slowly increasing.
  - Recent FDI commitments have potential to increase Bulgaria’s foothold in rapidly expanding GVCs, such as electric vehicles’ batteries.

### D. Room for Further GVC Integration with the EU
- Key findings on Bulgaria’s GVC integration
  - Bulgaria consistently lags peers in integration with other EU countries.
  - On the buying side:
    - Russian inputs amounted to 8 percent of Bulgaria’s backward linkages in 2020 compared to 12 percent for EU partners.
    - The geoeconomic fallout from Russia’s invasion of Ukraine is expected to weaken the economic relationship between the two countries.
  - On the selling side:
    - Bulgaria’s forward linkages with the EU are the lowest among peers, amounting to 12.2 percent of exports in 2020.
  - China’s role:
    - Although relatively limited, China’s importance as a buyer in Bulgaria’s GVCs has been steadily growing.
    - In 2020, China became the largest GVC partner destination for Bulgaria’s value added outside of the EU.
    - Dependence to China is particularly elevated in sectors such as basic metals and mining and quarrying of non-energy products.
  - Foreign Market Reliance (FMR):
    - FMR measures the ratio of domestic output used in foreign production to total domestic production and captures “total” exposure of domestic activity to downstream disruptions in GVCs.
    - FMR differs from forward linkages by considering exports consumed in destination countries and using gross trade flows (thus “double counts” value added that crosses borders multiple times), providing a more comprehensive picture of global chain risks.

### E. Drivers of GVC Integration (Empirical Results and ML)
- Common drivers considered
  - (i) land and natural resources endowments, labor and capital;
  - (ii) geographic position relative to large GVC hubs;
  - (iii) domestic industrial capacity;
  - (iv) openness to trade (tariffs and free trade agreements) and FDI;
  - (v) institutional quality;
  - (vi) transport connectivity;
  - (vii) macroeconomic factors such as cost competitiveness and degree of financial development.
- Machine learning (ML) panel regression over 2002–2020 main correlates of GVC integration:
  - Major drivers: high transport connectivity (density of road infrastructure), strong rule of law, and low wages (cost competitiveness).
  - Other important determinants: share of manufacturing in GDP (industrial structure), education of the labor force, investment in innovation (R&D), depth of financial markets, and internet connectivity (share of households using internet).
  - The analysis confirms deceleration in GVC integration during the Global Financial Crisis (GFC) and the COVID pandemic.
- Bulgaria’s position among EU peers:
  - Bulgaria belongs to a cluster of mostly small open EU economies characterized by relatively low wages, less developed infrastructure and financial intermediation, and predominantly low rule-of-law scores.
  - Bulgaria has the lowest real wage in the EU (cost advantage) but this advantage is eroding as wages catch up with the EU average.
  - Relative gaps where improvement is most needed: density of road infrastructure and rule of law; notable gaps also in R&D and internet connectivity.
  - Areas where Bulgaria fares better but still can improve: industrial intensity (share of manufacturing), share of labor force with a tertiary degree, and financial intermediation development—especially the quality of education.

### F. Policy Implications and Priorities
- Strategic shift
  - As cost advantage erodes, Bulgaria needs to move from cost to non-cost competitive advantages to increase participation in GVCs.
- Priority actions to improve non-price competitiveness and deepen GVC integration:
  - Step up investment in transport infrastructure (road density).
  - Strengthen governance and the rule of law.
  - Increase investment in R&D and enhance internet connectivity.
  - Improve skills and the quality of education to boost labor-force capabilities.
- Leveraging geopolitical opportunities:
  - Prospects of joining the euro and Schengen areas provide opportunities for stronger ties with regional EU countries.
  - A deeper integration with EU partners will require filling gaps with EU peers to become more competitive and attractive for foreign investors.
- Expected outcomes of policy actions
  - Increased GVC integration through higher backward and forward linkages.
  - Greater attraction of multinational firms and FDI, supporting moves up the value-added and technological ladder.

### G. ML Methods, Forecast Comparison, and Model Performance
- Data sources and transformations
  - GVC data from OECD TiVA dataset, 2023 vintage.
  - Drivers mostly from Eurostat (as of April 17, 2024), World Bank World Development Indicators (share of private credit in GDP), World Governance Indicators (institutional quality), and OECD transport infrastructure database (road density).
  - Variables centered and standardized; explanatory variables calculated as a percentage of their EU average values.
- ML methods applied
  - Linear Regression
  - Elastic Net
  - Support Vector Regression
  - Random Forest Regression
  - Extreme Gradient Boosting (XGBoost)
  - K-Nearest Neighbors
- Model training and validation
  - Dataset split: 75 percent training, 25 percent testing.
  - Hyperparameters set via cross-validated grid-search.
  - Model interpretability: SHAP values used to measure contribution of each driver to predictions.
- ML models forecasting performance (testing subsample)
  - Linear regression: MSE 0.07, R2 0.93
  - Elastic net: MSE 0.07, R2 0.93
  - k-Nearest Neighbors: MSE 0.05, R2 0.95
  - Support vector machine: MSE 0.05, R2 0.95
  - Random forest: MSE 0.06, R2 0.94
  - Extreme gradient boosting: MSE 0.09, R2 0.91
  - Overall: k-Nearest Neighbors, Support Vector Machine, and Random Forest perform slightly better than the rest.
- Modified Diebold-Mariano test (forecast comparison) key results
  - Positive sign indicates column model outperforms row model; significance: *p<0.1; **p<0.05; ***p<0.01
  - Selected pairwise statistics:
    - Linear regression vs Support vector machine: 2.44**
    - Linear regression vs k-Nearest Neighbors: 2.79***
    - Elastic net vs Support vector machine: 2.43**
    - Elastic net vs k-Nearest Neighbors: 2.72***
    - Support vector machine vs Extreme gradient boosting: -2.68***
    - Random forest vs Extreme gradient boosting: -2.34**
    - Extreme gradient boosting vs k-Nearest Neighbors: 2.53**
  - Key empirical findings from the test:
    - The SVM and kNN models have a statistically significant better forecasting performance than the other models.
    - The random forest performs better than the extreme gradient boosting model.
- SHAP analysis and feature importance
  - SHAP properties utilized: Efficiency, Symmetry, Dummy, Additivity.
  - Interpretation notes:
    - Distribution of SHAP values by feature is shown for each observation; wider spread implies higher variability and greater impact of the feature on the forecast.
    - SHAP values color-coded by original feature values: high values are blue, low values are red.
    - If a feature’s positive SHAP values are blue, higher feature values are associated with positive deviations of the output variable forecast from its expected value (example: rule of law).
    - If a feature’s positive SHAP values are red (example: real wages), this indicates a negative relationship between the feature and the output variable.
  - Feature importance across models:
    - Generally high contribution: road infrastructure, real wages, and rule of law.
    - Typically lower-ranking features: internet connectivity and the share of population with higher education.
    - Internet connectivity and higher-education-share impacts are less homogeneous across models.

### H. Cluster Analysis of Drivers of GVC Participation
- Clustering method and data
  - Ward’s method for hierarchical clustering minimizing total within-cluster variance based on Euclidean distance (Ward, 1963).
  - Averages for the 2010–20 period were used; data were preliminarily centralized and standardized.
  - Based on the within-group sum of squares, five clusters are identified.
- Variables (clusters of GVC correlates) used
  - Infrastructure: simple average of road infrastructure and internet connectivity scores.
  - Cost competitiveness: real wages.
  - Human capital and innovation: simple average of share of population with higher education and share of R&D in GDP.
  - Governance: rule-of-law indicator of the World Bank.
  - Financial depth: share of private credit in GDP.
  - Economic structure: share of manufacturing in GDP.
- Cluster characteristics and country groupings
  - Two broad country categories observed:
    - First group: large non-price competitiveness advantages (infrastructure, governance, human capital, innovation, financial depth).
    - Second group: high cost-competitiveness and higher share of manufacturing in GDP.
  - Bulgaria’s cluster (green): Bulgaria, Croatia, Estonia, Italy, Latvia, and Lithuania — characterized by low wages and low levels of infrastructure, human capital and innovation, governance, and financial integration, while being relatively manufacturing-oriented.
  - Purple cluster: Czechia, Hungary, Poland, Romania, Slovakia, and Slovenia — similar to green but with higher share of manufacturing in GDP and somewhat lower level of highly educated population.
  - Orange cluster: Portugal and Spain — subpar infrastructure and lower wages, but high rule-of-law scores and depth of financial markets.
  - Blue cluster: Belgium, Denmark, France, Luxembourg, Malta, and Netherlands — somewhat more developed infrastructure and lower industrialization relative to comparable groups.
- Visual outputs referenced (in source)
  - Dendrogram showing the five clusters and hierarchical relationships.
  - Figures showing clusters in terms of variable pairs (each cluster represented by a different color).

*Source: sipea2024023 — IMF staff calculations.*

### REFERENCES

### sipea2024023 - REFERENCES

### A. Introduction and Main Messages
- Bulgaria is a small open economy whose integration in global value chains (GVCs) grew before the global financial crisis (GFC) and EU accession, but has since grown only modestly and remains below peers in links with EU countries.
- To harness opportunities from joining the euro and Schengen areas, Bulgaria should improve non-cost competitiveness by improving governance and investing in infrastructure and human capital.
- Authors: Prepared by Giacomo Magistretti and Iglika Vassileva (both EUR).

### B. Developments in Foreign Trade and FDI
- Trade dynamics
  - Bulgaria’s share in world exports steadily increased over the last 3 decades.
  - Exports to the EU rose from 15 percent of GDP in 1999 to 25 percent in 2007, then grew more gradually in the last decade.
  - Notable recent increases in exports to Germany and Romania; declines in importance of Italy, Greece, and Türkiye as export destinations.
  - Imports from the EU have decreased lately, while China, Türkiye, Serbia, and the US increased their importance as suppliers.
  - Exports are shifting toward more technologically-intensive products, with rises in electrical and other machinery and chemicals and declines in textiles and apparel; however, the Economic Complexity Index of exports still lags peers and has not increased over the last two decades.
- Geopolitical energy shifts
  - Import of Russian gas significantly declined after Russia stopped pipeline gas supply to Bulgaria in April 2022.
  - Import of Russian crude oil increased under an EU derogation; a more notable decline in reliance on Russian oil was expected only starting from 2024 after the Bulgarian parliament revoked the derogation from March 1, 2024.
- Foreign Direct Investment (FDI)
  - FDI remains substantially below its 2007 peak and settled to levels similar to other newer member states (NMSs) in the last decade.
  - Recent FDI largely came from EU partners and was mostly concentrated in services, especially financial activities (including real estate), and wholesale and retail trade.
  - Making Bulgaria more attractive to foreign investors is expected to boost integration and attract multinational firms relocating in Europe.

### C. Bulgaria’s Participation in Global Value Chains (GVCs)
- Overall participation and direction
  - In 2020, 35 percent of Bulgaria’s final demand was met by value added coming from abroad, about half of which was from other EU countries.
  - Bulgaria sends a sizeable share of its value added abroad: value added going abroad peaked at almost 45 percent in 2017 and declined to about 38 percent in 2020, close to the EU average.
  - Almost half of value added sent abroad went to other EU countries; the share going to China grew steadily and China became the largest destination outside the EU in 2020, overtaking the US.
- Sectoral composition of value added sent abroad (2020)
  - 25 percent from manufacturing (including construction and utilities).
  - 13 percent from agriculture and mining.
  - 56 percent from services.
  - Shares of wholesale and retail trade and IT and communication in value added sent abroad have been growing steadily over time.
- Labor market interconnectedness
  - Fraction of Bulgarian employment sustained by foreign demand rose from 35 percent during the GFC to about 50 percent by the end of the 2010s, then declined slightly to 46 percent in 2020.
  - Bulgaria’s employment sustained by foreign demand is the fourth highest in the EU behind Ireland, Luxembourg, and Malta.
- Aggregate GVC metrics and trends
  - Bulgaria’s participation in GVCs increased strongly before EU accession and the GFC, declined in 2009, partially recovered, and then increased only marginally thereafter.
  - Bulgaria’s overall GVC participation trails most NMSs and the EU average.
  - Backward linkages (foreign value added embodied in domestic exports) are in line with peers and the EU average.
  - Forward linkages (domestic value added embodied in foreign exports) are limited, below the EU average and below peer countries, though slightly increasing in recent years.
- Industry-level linkages
  - Backward linkages: the largest involvement is in basic metals and coke and refined petrol products, accounting for almost a third of total foreign value added in Bulgaria’s exports in 2020.
  - Forward linkages: well-balanced across manufacturing and services industries.
- Sophistication
  - Involvement in information GVCs (computers, electronics, communication, IT, and information services) is among the lowest in the EU, though sophistication has been slowly increasing.
  - Recent FDI commitments have potential to increase Bulgaria’s foothold in rapidly expanding GVCs, such as electric vehicles’ batteries.

### D. Determinants and Policy Implications
- Key policy priorities identified
  - Improve non-price competitiveness by strengthening governance.
  - Invest in infrastructure.
  - Invest in human capital and innovation.
  - Create a more attractive economic environment for foreign investors, including those relocating within Europe amid geoeconomic fragmentation.
- Expected outcomes of policy actions
  - Increased GVC integration through higher backward and forward linkages.
  - Greater attraction of multinational firms and FDI, supporting moves up the value-added and technological ladder.

*Source: BULGARIA IN GLOBAL VALUE CHAINS: LEVERAGING INTEGRATION WITH THE EU, INTERNATIONAL MONETARY FUND, May 10, 2024.*

### 16.      There is room for further increase in GVC integration with the EU, on both the selling

### 16.      There is room for further increase in GVC integration with the EU, on both the selling and, especially, the buying side.

### Key findings on Bulgaria’s GVC integration
- Bulgaria consistently lags peers in integration with other EU countries.
- On the buying side:
  - Russian inputs amounted to 8 percent of Bulgaria’s backward linkages in 2020 compared to 12 percent for EU partners.
  - The geoeconomic fallout from Russia’s invasion of Ukraine is expected to weaken the economic relationship between the two countries.
- On the selling side:
  - Bulgaria’s forward linkages with the EU are the lowest among peers, amounting to 12.2 percent of exports in 2020.
- China’s role:
  - Although relatively limited, China’s importance as a buyer in Bulgaria’s GVCs has been steadily growing.
  - In 2020, China became the largest GVC partner destination for Bulgaria’s value added outside of the EU.
  - Dependence to China is particularly elevated in sectors such as basic metals and mining and quarrying of non-energy products.
- Foreign Market Reliance (FMR):
  - FMR measures the ratio of domestic output used in foreign production to total domestic production and captures “total” exposure of domestic activity to downstream disruptions in GVCs.
  - FMR differs from forward linkages by considering exports consumed in destination countries and using gross trade flows (thus “double counts” value added that crosses borders multiple times), providing a more comprehensive picture of global chain risks.

### Drivers of GVC integration (empirical results)
- Common drivers considered: (i) land and natural resources endowments, labor and capital; (ii) geographic position relative to large GVC hubs; (iii) domestic industrial capacity; (iv) openness to trade (tariffs and free trade agreements) and FDI; (v) institutional quality; (vi) transport connectivity; (vii) macroeconomic factors such as cost competitiveness and degree of financial development.
- Machine learning (ML) panel regression over 2002–2020 identifies main correlates of GVC integration:
  - Major drivers: high transport connectivity (density of road infrastructure), strong rule of law, and low wages (cost competitiveness).
  - Other important determinants: share of manufacturing in GDP (industrial structure), education of the labor force, investment in innovation (R&D), depth of financial markets, and internet connectivity (share of households using internet).
  - The analysis confirms deceleration in GVC integration during the Global Financial Crisis (GFC) and the COVID pandemic.
- Bulgaria’s position among EU peers:
  - Bulgaria belongs to a cluster of mostly small open EU economies characterized by relatively low wages, less developed infrastructure and financial intermediation, and predominantly low rule-of-law scores.
  - Bulgaria has the lowest real wage in the EU (cost advantage) but this advantage is eroding as wages catch up with the EU average.
  - Relative gaps where improvement is most needed: density of road infrastructure and rule of law; notable gaps also in R&D and internet connectivity.
  - Areas where Bulgaria fares better but still can improve: industrial intensity (share of manufacturing), share of labor force with a tertiary degree, and financial intermediation development—especially the quality of education.

### Policy implications and priorities
- As cost advantage erodes, Bulgaria needs to move from cost to non-cost competitive advantages to increase participation in GVCs.
- Priority actions to improve non-price competitiveness and deepen GVC integration:
  - Step up investment in transport infrastructure (road density).
  - Strengthen governance and the rule of law.
  - Increase investment in R&D and enhance internet connectivity.
  - Improve skills and the quality of education to boost labor-force capabilities.
- Leveraging geopolitical opportunities:
  - Prospects of joining the euro and Schengen areas provide opportunities for stronger ties with regional EU countries.
  - A deeper integration with EU partners will require filling gaps with EU peers to become more competitive and attractive for foreign investors.

### Machine learning approach and model details
- Data sources:
  - GVC data from OECD TiVA dataset, 2023 vintage.
  - Drivers mostly from Eurostat (as of April 17, 2024), World Bank World Development Indicators (share of private credit in GDP), World Governance Indicators (institutional quality), and OECD transport infrastructure database (road density).
- Data transformations:
  - Variables centered and standardized; explanatory variables calculated as a percentage of their EU average values.
- ML methods applied:
  - Linear Regression
  - Elastic Net
  - Support Vector Regression
  - Random Forest Regression
  - Extreme Gradient Boosting (XGBoost)
  - K-Nearest Neighbors
- Model training and validation:
  - Dataset split: 75 percent training, 25 percent testing.
  - Hyperparameters set via cross-validated grid-search.
  - Model interpretability: SHAP values used to measure contribution of each driver to predictions.
- SHAP explanation summary:
  - SHAP (Kernel SHapley Additive exPlanations) approximates Shapley values to estimate each feature’s expected marginal contribution to deviations in the outcome prediction.
  - SHAP properties utilized: Efficiency, Symmetry, Dummy, Additivity.
- ML models forecasting performance (testing subsample):
  - Linear regression: MSE 0.07, R2 0.93
  - Elastic net: MSE 0.07, R2 0.93
  - k-Nearest Neighbors: MSE 0.05, R2 0.95
  - Support vector machine: MSE 0.05, R2 0.95
  - Random forest: MSE 0.06, R2 0.94
  - Extreme gradient boosting: MSE 0.09, R2 0.91
  - Overall: k-Nearest Neighbors, Support Vector Machine, and Random Forest perform slightly better than the rest.

*Sources: OECD TiVA 2023 and IMF staff calculations.*

### 2.      The modified Diebold-Mariano test for forecast comparison

### 2.      The modified Diebold-Mariano test for forecast comparison

### Modified Diebold-Mariano test implementation
- The (Harvey, Leybourne, & Newbold., 1997) modification of the Diebold-Mariano (1995) test is implemented to improve finite sample properties by correcting almost entirely the bias of the Diebold-Mariano test, yielding an approximately unbiased estimate of variance of loss differential. (Footnote text preserved.)

### Forecast comparison results (Table 2)
- Interpretation note: A positive sign of the statistics indicates that the model in the column performs better than the model in the row.
- Note on significance: *p<0.1; **p<0.05; ***p<0.01
- Pairwise modified Diebold-Mariano statistics (row vs column; positive means column model outperforms row model):
  - Linear regression vs:
    - Elastic net: 0.21
    - Support vector machine: 2.44**
    - Random forest: 0.73
    - Extreme gradient boosting: -1.35
    - k-Nearest Neighbors: 2.79***
  - Elastic net vs:
    - Support vector machine: 2.43**
    - Random forest: 0.71
    - Extreme gradient boosting: -1.35
    - k-Nearest Neighbors: 2.72***
  - Support vector machine vs:
    - Random forest: -1.02
    - Extreme gradient boosting: -2.68***
    - k-Nearest Neighbors: -0.5
  - Random forest vs:
    - Extreme gradient boosting: -2.34**
    - k-Nearest Neighbors: 0.71
  - Extreme gradient boosting vs:
    - k-Nearest Neighbors: 2.53**
- Key empirical findings from the test:
  - The SVM and kNN models have a statistically significant better forecasting performance than the other models (as shown in Table 2).
  - The random forest performs better than the extreme gradient boosting model.

### Overall ML model performance and SHAP analysis
- Overall assessment:
  - Despite some differences in forecasting performance, ML models perform similarly; therefore results are cast in terms of average SHAP values.
- SHAP value interpretation and visualization notes:
  - Distribution of SHAP values by feature is shown for each observation; wider spread implies higher variability in SHAP values and greater impact of the feature on the forecast of GVC integration.
  - SHAP values are color-coded by original feature values: high values are blue, low values are red.
  - If a feature’s positive SHAP values (to the right of the vertical line) are blue, higher feature values are associated with positive deviations of the output variable forecast from its expected value (example given: rule of law).
  - If a feature’s positive SHAP values are red (example given: real wages), this indicates a negative relationship between the feature and the output variable.
- Feature importance across models:
  - Generally high contribution: road infrastructure, real wages, and rule of law.
  - Typically lower-ranking features: internet connectivity and the share of population with higher education.
  - Internet connectivity and higher-education-share impacts are less homogeneous across models.

### Cluster analysis of the drivers of GVC participation
- Clustering method and data:
  - Ward’s method for hierarchical clustering is used, minimizing total within-cluster variance based on Euclidean distance (Ward, 1963).
  - Averages for the 2010–20 period were used; data were preliminarily centralized and standardized.
  - Based on the within-group sum of squares, five clusters are identified.
- Variables (clusters of GVC correlates) used in the cluster analysis:
  - Infrastructure: simple average of road infrastructure and internet connectivity scores.
  - Cost competitiveness: real wages.
  - Human capital and innovation: simple average of share of population with higher education and share of R&D in GDP.
  - Governance: rule-of-law indicator of the World Bank.
  - Financial depth: share of private credit in GDP.
  - Economic structure: share of manufacturing in GDP.
- Cluster characteristics and country groupings:
  - Two broad country categories observed:
    - First group: large non-price competitiveness advantages (infrastructure, governance, human capital, innovation, financial depth).
    - Second group: high cost-competitiveness and higher share of manufacturing in GDP.
  - Bulgaria’s cluster (green): Bulgaria, Croatia, Estonia, Italy, Latvia, and Lithuania — characterized by low wages and low levels of infrastructure, human capital and innovation, governance, and financial integration, while being relatively manufacturing-oriented.
  - Purple cluster: Czechia, Hungary, Poland, Romania, Slovakia, and Slovenia — similar to green but with higher share of manufacturing in GDP and somewhat lower level of highly educated population.
  - Orange cluster: Portugal and Spain — subpar infrastructure and lower wages, but high rule-of-law scores and depth of financial markets.
  - Blue cluster: Belgium, Denmark, France, Luxembourg, Malta, and Netherlands — somewhat more developed infrastructure and lower industrialization relative to comparable groups.
- Visual outputs referenced:
  - Dendrogram showing the five clusters and hierarchical relationships.
  - Figures showing clusters in terms of variable pairs (each cluster represented by a different color).

*Source: sipea2024023 — IMF staff calculations.*

---


_Source: https://www.imf.org/-/media/files/publications/selected-issues-papers/2024/english/sipea2024023.pdf_