Recessions forecasting is invaluable to policymakers, economists, and investors because it enables planning, decision-

making, and risk mitigation. This paper suggests a hybrid approach that merges ensemble machine learning (ML),

unsupervised clustering, and Response Surface Methodology (RSM) to expand predictive accuracy and

interpretability. Employing U.S. quarterly data (1966–2025) of six macroeconomic variables, the methodology

incorporates feature selection, SHAP importance analysis, Bayesian hyperparameter tuning, and RSM estimation.

Unsupervised techniques (DBSCAN, Gaussian Mixture Models) revealed hidden structures consistent with supervised

predictions, and XGBoost provided the greatest ROC-AUC. Though RSM resulted in lower accuracy, it produced

explicit functional relations. Collectively, these techniques show that hybrid modeling provides predictive capability

along with interpretability and that their integration is useful to make reliable recession forecasts.