Comparing machine learning and classical methods in forecasting Türkiye’s macroeconomic performance

Abstract

This study forecasts Türkiye’s medium-term macroeconomic performance through a composite index based on growth, unemployment, inflation, the budget balance, and the current account balance. Quarterly data for 2006Q1–2026Q1 are used to compare artificial neural networks, ARIMA/SARIMA models, and ordinary least squares regression. Model performance is evaluated through rolling-origin validation over 40 out-of-sample periods using MAE, MAPE, and RMSE. The statistical significance of forecast error differences is examined with the Diebold–Mariano test, while the sensitivity of the neural network is assessed across 180 hyperparameter configurations. The results show that the neural network produces the lowest error in the baseline specification, although its advantage is not statistically significant at the 5 percent level. When seasonal information and lagged component values are included, OLS yields the lowest forecast error. Conditional forecasts from the preferred OLS specification place the index between 91.66 and 94.56 during 2026Q2–2028Q1. The projected path remains broadly stable, with seasonal fluctuations but no pronounced upward or downward trend. Overall, the findings indicate that forecast performance depends on model specification and that claims of machine learning superiority should be evaluated cautiously.