Performance Comparison of Machine Learning-Based Regression Models in Predicting Türkiye’s Exports

Abstract

The study aims to identify the most suitable machine learning-based regression model that can be used to predict the future course of exports by analysing the macroeconomic indicators affecting Türkiye’s export performance and through variables with a high correlation with exports. The study utilises 12 different macroeconomic variables related to the Turkish and global economies, and the relationship between these variables and exports are examined using correlation analysis. For the variables showing a strong relationship with exports, the linear model LASSO is applied alongside four different ensemble-based machine learning algorithms, using the Python programming language. To obtain unbiased and robust predictive accuracy, model performances are evaluated using a cross-validation framework. The results demonstrate that all models achieve high predictive performance; however, LASSO, in particular, provides the strongest overall accuracy with near-perfect explanatory power and a low error rate. Among the ensemble methods, Gradient Boosting emerges as the approach with the best performance. The findings demonstrate that regression methods can effectively capture Türkiye’s export dynamics and provide a reliable foundation for export forecasts. The study contributes to the literature from both methodological and practical perspectives in understanding Türkiye’s export dynamics, whilst also supporting data-driven policy development for decision-makers.