Forecasting Turkish exports: A comparative study of statistical, machine learning, and deep learning methods

Abstract

The purpose of this study is to identify the most accurate forecasting approach for Turkish monthly merchandise exports by rigorously comparing statistical, machine learning, and deep learning methods on a low-frequency macroeconomic series characterized by structural breaks and regime changes. Regarding methodology, eight models are evaluated, SARIMAX, Prophet, Hybrid-S (SARIMAX+CatBoost), Hybrid-P (Prophet+CatBoost), XGBoost, CatBoost, Gated Recurrent Unit (GRU), and Neural Basis Expansion Analysis for Time Series (N-BEATS), applied to monthly Turkish export data spanning January 2002 to December 2025 (288 observations), evaluated over an 83-month walk-forward test period (February 2019– December 2025), with Brent crude oil price, real effective exchange rate, and industrial production index as exogenous variables. Walk-forward expanding-window validation eliminates look-ahead bias, hyperparameters are optimized via Optuna with 50 trials, and statistical significance is assessed using Diebold-Mariano tests. The findings reveal that the SARIMAX+CatBoost hybrid achieves the highest predictive accuracy (R² = 0.857, MAPE = 5.67%), marginally outperforming standalone SARIMAX (R² = 0.853, MAPE = 5.72%), while deep learning methods (GRU, N-BEATS) perform substantially below all other models (R² = 0.44 and 0.39, respectively), consistent with the literature on neural networks applied to low-frequency, small-sample macroeconomic series. In terms of contribution, this study presents the first unified eight-model benchmark for Turkish export forecasting under a strictly leakage-free walk-forward validation protocol, introduces two novel hybrid architectures validated across multiple structural regimes including the COVID-19 shock, and empirically confirms the data-regime hypothesis: statistical and hybrid methods dominate in low-frequency, small-sample, structurally unstable macroeconomic environments.