This study develops a hybrid econometric and deep learning framework for forecasting global commodity price dynamics using multi-commodity time series data over the period 2000-2026. The proposed framework integrates Autoregressive Integrated Moving Average (ARIMA) models to capture linear dependencies, Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models to model volatility persistence, and machine learning techniques, including Random Forest and Extreme Gradient Boosting, to extract nonlinear structures from residual components. In addition, Long Short-Term Memory (LSTM) networks are employed to capture long-term temporal dependencies. Empirical results across major commodities such as crude oil, gold, copper, natural gas, platinum, and silver demonstrate that the hybrid model significantly outperforms standalone econometric and deep learning models. The framework achieves consistent reductions in forecasting errors, with improvements in predictive accuracy ranging from approximately 5.5% to 12% across commodities and generates positive out-of-sample R² values in rolling-window evaluations, indicating strong temporal stability. The results further reveal pronounced volatility persistence (α + β ≈ 0.98-0.99) across all commodities and confirm that nonlinear residual components contain substantial predictive information. While LSTM models capture general temporal patterns, their standalone predictive performance remains limited, with relatively low or negative R² values in several cases. The findings indicate that commodity price dynamics are inherently multidimensional, characterized by linear dependencies, volatility clustering, and nonlinear interactions. By integrating these components within a unified structure, the proposed hybrid framework provides a robust and scalable approach for commodity price forecasting, with important implications for financial modeling, investment strategies, and policy analysis.