Abstract
Machine learning was applied to predict Arabica and Robusta coffee prices 2-6 months ahead using climatic, production, and economic data from 2009 to 2025. SHAP analysis revealed that global and Brazilian stock levels, Vietnamese drought, Colombian rainfall, and U.S. dollar exchange rate were the most influential drivers of price variation, showing that even features with weak simple correlations can have high predictive power. We compared Multilayer Perceptron (MLP), Extreme Learning Machine (ELM), Random Forest (RF), XGBoost, and two stacking configurations (RF/XGBoost/ELM) and (RF/XGBoost), both combined through a meta-learner. For Robusta, the three-model stacking configuration (RF/XGBoost/ELM) achieved the best overall accuracy with Pearson's correlation coefficient (PA) of 0.9264, Hit Rate (HR) of 92.8000%, Root-Mean-Square Error (RMSE) of 0.0793, and Mean Absolute Percentage Error (MAPE) of 9.3051%. Arabica proved more difficult to forecast, and the ELM delivered the highest independent test performance with PA of 0.9064 and HR of 94.1167%. To assess model robustness under future warming, we also tested two extreme climate scenarios derived from historical data and literature for major Arabica and Robusta producing regions (Brazil, Colombia, and Vietnam). These scenarios combined higher temperatures, reduced production, lower stocks, expanded production area, increased global consumption, and concurrent logistic and climatic shocks. Despite these stringent assumptions, the models did not project strong price surges and showed greater difficulty when extrapolating Arabica prices, while Robusta forecasts remained more consistent with realistic market values. These results underscore the ability of machine-learning models and SHAP interpretation to reveal the complex economic, climatic and productive factors governing global coffee price dynamics and highlight their conservative extrapolation when confronted with unprecedented climatic conditions.