Free access: https://rdcu.be/e2dIY
This article builds on more comprehensive research that examines the patterns and determinants of SC, as shown in Salinas (2025) and Salinas & Zhang (2025). It is among the first to systematically apply machine learning within the system approach to structural change, providing data-driven insights into countries' structural tendencies.
The article hypothesizes that certain shared variables may influence the likelihood of triggering or constraining countries' structural tendencies, thereby shaping the CIP's midterm trend.
This hypothesis is addressed by solving a classification problem: predicting CIP trends corresponding to positive (1) and negative changes (0) for each country and midterm period between 2000 and 2021 (2000–2006, 2007–2013, and 2014–2021).
How are the determinants of the CIP midterm trends unveiled? Unlike traditional econometrics, which relies on prior model specification and in-sample inference, this study employs supervised machine learning for out-of-sample inference using a representative sample of 148 economies. Features exhibiting strong predictive power are framed as key determinants that can trigger cumulative causation mechanisms, shaping countries' structural tendencies. This approach substantiates out-of-sample inference within the system approach to structural change, evidencing the relevance of features in shaping structural tendencies across a range of new scenarios and samples.
This application establishes machine learning as an epistemic lens for analyzing data patterns. In this vein, explainable AI techniques such as SHAP (SHapley Additive exPlanations) enhance the interpretability of parametric and nonparametric algorithms, facilitating the analysis of feature importance and dependence in predicting countries' CIP trends. Therefore, this methodological approach demonstrates the potential of supervised ML to reveal complex nonlinear relationships, providing robust analytical foundations for data-driven policies that foster structural change in nations.
Thus, this article identifies determinants as those features that meet two criteria: (i) features with the potential to engender cumulative causation mechanisms shaping countries' CIP trends, and (ii) features that demonstrate strong predictive performance across a range of scenarios and samples.
Hence, the system approach to structural change underpins the potential of features to unleash cumulative and feedback effects that engender systemic transformations and shape countries' structural tendencies, thereby influencing CIP midterm trends. At the same time, machine learning substantiates the generalizability of features across diverse scenarios and samples.
Among the 61 evaluated features, ten determinants demonstrated noteworthy out-of-sample predictive performance for CIP trends:
- Midterm growth of exports of merchandise at current prices
- Midterm growth of the human development index
- Midterm growth of imports of merchandise at current prices
- Midterm growth of industry at constant prices
- Income elasticity of demand for exports
- Midterm growth of agriculture at constant prices
- Midterm growth of the export product diversification index
- Midterm growth of the terms of trade trend
- Midterm growth of the institutional dimension of the productive capacities index (PCI)
- Midterm growth of the information and communication technology dimension of PCI
Section 1 introduces the study. Section 2 examines the prediction category, the supervised ML analytic method, the relationship between the CIP and the system approach to SC, and the relevance of supervised ML in uncovering the triggers of the CIP midterm trend. Section 3 outlines the methodological approach, including the data and variables studied, the algorithms employed, and the metrics used to evaluate their performance. Section 4 presents the research findings and implications. Finally, Section 5 provides concluding observations and reflections.
Article highlights:
- Unveils the dynamic determinants of the CIP midterm trend for a representative sample of economies.
- Demonstrates the relevance of the system approach to SC in underpinning the role of predictive features in shaping the CIP midterm trend.
- Sheds light on the potential of supervised ML to address real-world problems by demonstrating the generalizability of features and algorithms.
- Provides actionable insights with theoretical and policy implications for the field of structural change.
Salinas, J. (2025). Unveiling structural change patterns: An unsupervised machine learning approach to long-term dynamics. Humanities and Social Sciences Communications, 12(1). https://doi.org/10.1057/s41599-025-05128-9
Salinas, J., & Zhang, J. (2025). Unveiling Structural Change Determinants: A Machine Learning Approach to Long-Term Dynamics. Socio-Economic Planning Sciences, 101, 102290. https://doi.org/10.1016/j.seps.2025.102290