From Feature Selection to Forecasting: A Two-Stage Hybrid Framework for Food Price Prediction Using Economic Indicators in Turkey

This study develops a comprehensive two-stage hybrid framework to forecast food prices in Turkey, addressing inflation prediction challenges critical for sustainable food security in emerging economies. In the first stage, systematic relationship and causality analyses—comprising correlation, ARDL, cointegration, and Granger causality tests—identified ten key predictors from the Turkish Statistical Institute and Central Bank datasets. In the second stage, ten predictive models, including ensemble (Gradi-ent Boosting, Random Forest, SVR), traditional (ARIMA, Linear Regression), and deep learning approaches (LSTM, NARX-RNN, ANFIS), were evaluated using rice prices as a pilot case. Ensemble models demonstrated clear superiority, with Gradient Boosting achieving optimal single-split performance (R² = 0.9990) and high cross-validation consistency (mean R² = 0.9742 ± 0.03). Support Vector Regression (R² = 0.9896 ± 0.02) and Random Forest (R² = 0.9811 ± 0.02) showed statistically equivalent performance, reinforcing ensemble robustness. NARX-RNN analysis revealed a six-month lag in economic shock transmission, providing a practical policy intervention window. SHAP-based interpretability identified insurance, healthcare, transportation, educa-tion, and social protection expenditures as major drivers of food price formation, un-derscoring Turkey's cross-sector inflation mechanisms. These findings integrate econometric rigor with machine learning transparency, offering practical tools for sustainable inflation management and early warning systems in volatile emerging markets.