Food insecurity is a complex challenge influenced by climatic, economic, agricultural, conflict-related, and socio-demographic factors. Data-driven decision-making is essential for designing targeted interventions and improving policies. This study presents spatiotemporal integration of heterogeneous datasets incorporating climate, crop production, food price, conflict, cross-border trade, exchange rate, and population growth for predicting food security in Ethiopia. All datasets were aligned using consistent region–month key and converted to a monthly resolution to preserve temporal relationships across regions. New variables, Food Price Index, Conflict Exposure Index, Climatic Vulnerability Index, and Market Accessibility Index, are derived through feature engineering and then included for predictive model development. Food Security Index (FSI) is the target variable, and preliminary results show that food insecurity is increasing in Ethiopia with distinct variations across regions. We used the stacking ensemble of statistical models, i.e., Random Forest (RF) and Extreme Gradient Boosting (XGBoost), and deep learning models, i.e., Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN), using LSTM as a Meta-learner. A strict time-based split (earlier months for training and later months for testing) is applied to prevent temporal information leakage. The proposed stacking model outperforms individual learners with smaller errors, i.e., MAE = 0.00201 and RMSE = 0.038, and a higher R², i.e., 0.96. This study proves that spatiotemporal integration of heterogeneous datasets with stacking ensemble model provides holistic and accurate prediction output. This approach is valuable and recommendable for policymakers and early warning systems for decision-making and policy improvement, over using statistical methods alone.