Forecasting Economic Growth and GDP per Capita in Syria Using XGBoost

Abstract

This study aims to forecast economic growth and GDP per capita growth using the XGBoost model within an annual time-series framework for the period 1961–2023. The analysis begins with descriptive statistics and visual inspection of both series, followed by diagnosing temporal dependence through the autocorrelation and partial autocorrelation functions to motivate the use of multiple lagged features. The data are then reformulated as a supervised learning problem by constructing a time trend and shared lag structures for both variables. A strict time-based split is applied, with 1961–2018 used for training and 2019–2023 for testing. Model hyperparameters are tuned via time-series cross-validation and randomized search, while early stopping is employed to mitigate overfitting. The results show that XGBoost substantially outperforms a simple benchmark model in forecasting accuracy for both economic growth and GDP per capita growth, as evidenced by improvements in mean squared error, root mean squared error, mean absolute percentage error, and the coefficient of determination, with statistically significant differences in predictive accuracy. Finally, the study produces forecasts through 2030 and reports uncertainty intervals obtained from residual-based simulation, supporting scenario-based planning and policy analysis in the Syrian context.