The validity of Rodrik's conclusion on real exchange rate and economic growth: factor priority evidence from feature selection approach

The study examines the relative importance (priority) of the real exchange rate—particularly its undervaluation—versus GDP per capita in driving economic growth, extending Rodrik’s (2008) claims using nonparametric machine learning methods. While prior work shows that poor exchange rate policy can harm growth, little is known about the priority of exchange rate measures among determinants of growth, and whether Rodrik’s conclusion holds across different functional forms (linear vs nonlinear) and economic contexts (developed, emerging, and developing economies). The paper evaluates these questions for Germany (developed), Slovakia (developed but highly open), and South Africa (emerging), emphasizing the policy relevance of identifying which factor should be prioritized to achieve sustainable growth.

The paper reviews two strands: (1) effects of exchange rate regimes on growth (e.g., Levy-Yeyati and Sturzenegger, 2003; Harms and Kretschmann, 2009; Sokolov et al., 2011; Ma and McCauley, 2011; De Vita and Kyaw, 2011; Benhima, 2012); and (2) links between real exchange rate (RER) levels/misalignment and growth (e.g., Tharakan, 1999; Rodrik, 2008; Tarawalie, 2010; Abida, 2011; Elbadawi et al., 2012; Glüzmann et al., 2012; Vieira et al., 2013). Easterly (2003) argues overvaluation harms growth, often amid macroeconomic imbalance and instability. Rodrik (2008) finds that RER undervaluation stimulates growth in developing countries, robust to various RER measures and estimators. However, prior studies mainly use parametric methods and do not assess the priority of exchange rate measures among growth determinants nor explore model functional forms; no prior work in economics has applied the specific machine learning techniques used here to this question.

Data and variables: Annual time-series for Germany, South Africa, and Slovakia, 1990–2016, sourced from World Bank (online). Variables follow Rodrik’s model: GDP growth (dependent variable), initial income per capita (RGDP per capita), and an index of real exchange rate undervaluation (UN). All variables are log-transformed to harmonize units and mitigate heteroscedasticity. A Ramsey RESET (LM) test indicates nonlinearity, motivating nonparametric methods. Rodrik model: Growth_it = α + β ln RGDP_it−1 + δ ln UN_it + f_i + f_t + μ_it. The undervaluation index uses the Balassa–Samuelson framework: In UN_i = In RER − In RER_it, with predicted In RER = a + β ln RGDP_it + f_e + M_it. Machine learning approach: Three techniques are employed to both model nonlinear relationships and assess variable priority. - Feature selection (FS) via multilayer artificial neural network (ANN): Multilayer ANN with hidden layers processes inputs (e.g., GDP per capita, RER undervaluation) to predict growth. FS aims to identify the most relevant predictors, reducing estimation cost and improving interpretability. Weight updates use the generalized delta rule with learning rate η. - Particle Swarm Optimization (PSO): A population-based stochastic optimizer updates particle velocities and positions using inertia and cognitive/social components to search for optimal model parameters and assess variable priority through permutations/combinations. Advantages include ease of implementation, computational efficiency, and robust performance in continuous nonlinear optimization. - Genetic Algorithm (GA): An evolutionary optimizer employing selection, crossover, and mutation to search parameter space; member selection is fitness-proportional (P_i = f_i / Σ f_k). GA has broad use in engineering and energy demand forecasting; here it is used to optimize model weights and compare against PSO for improving the Rodrik model. Model comparison: PSO and GA are each run with 1000 iterations to identify the better-performing optimization framework per country (based on error statistics), and FS is used to determine the most influential predictor of growth within the Rodrik specification.

Model performance (1000 iterations): - Germany: PSO outperforms GA (Error St.D.: PSO 0.084636 vs GA 0.15645), indicating PSO provides a better improvement to the Rodrik model for a developed economy like Germany. - South Africa: GA outperforms PSO (Error St.D.: GA 0.081297 vs PSO 0.09148), suggesting GA yields a more reliable model for the emerging economy. - Slovakia: PSO outperforms GA (Error St.D.: PSO 0.81139 vs GA 1.4759), indicating PSO is preferable for Slovakia. Feature selection (priority of determinants): - Germany: GDP per capita is the most relevant determinant (coefficient estimate ≈ 0.34, standard error ≈ 0.009), positively and significantly associated with growth; undervaluation is not significant. - South Africa: GDP per capita is also the key determinant (coefficient ≈ 0.44, standard error ≈ 0.007), with undervaluation not significant. - Slovakia: Real exchange rate undervaluation is the priority determinant (estimate ≈ 0.47; reported cost 1.02); GDP per capita is not the dominant factor. Comparative context: Slovakia’s high trade openness (~190% of GDP in 2018) and strong merchandise trade growth (11.6%) coincide with undervaluation’s primacy for growth. Germany’s openness is lower (87% of GDP; 7.8% merchandise trade growth; 1.4% GDP growth in 2018). South Africa shows relatively low openness (~60% of GDP; 5.6% merchandise trade growth; 0.8% GDP growth), aligning with GDP per capita, not undervaluation, as the main growth determinant.

The findings address whether the exchange rate (undervaluation) should be prioritized among growth determinants and whether Rodrik’s conclusion generalizes across contexts and nonlinear specifications. Results suggest conditional validity: in highly open economies (e.g., Slovakia), undervaluation takes precedence and is directly associated with higher growth, consistent with the tradables-sector transmission channel emphasized by Rodrik. In less open or differently structured economies (Germany and South Africa), GDP per capita is the dominant determinant and undervaluation does not significantly affect growth, indicating that exchange-rate-focused policies may be less effective where openness and tradables-sector dynamics are weaker. The machine learning results thereby refine Rodrik’s conclusion by highlighting trade openness and economic structure as key moderators of the exchange rate–growth relationship.

Using nonparametric machine learning (feature selection via ANN, PSO, and GA) on 1990–2016 data for Germany, South Africa, and Slovakia, the study finds that undervaluation stimulates growth only in Slovakia, a highly open economy, while GDP per capita is the most important determinant in Germany and South Africa. Thus, the priority of the real exchange rate as a growth determinant is context-dependent: it takes precedence in economies with high trade openness, whereas income per capita dominates where openness is lower. Policy recommendations follow: highly open economies should focus on viable exchange rate policies (e.g., currency undervaluation) to support sustained growth, while relatively less open economies should prioritize policies that raise income per capita over exchange rate interventions.

The analysis covers only three countries over 1990–2016, which may limit generalizability. The authors note a data problem for Slovakia affecting model fitting in part of the analysis. Results rely on feature selection and optimization within a specific ML framework and variable set (Rodrik’s specification), which may not capture all relevant determinants or country-specific factors.