HybridStack: Explainable hybrid ensemble for accurate inflation forecasting with high-frequency macro-financial data

Inflation forecasting plays a critical role in macroeconomic policy and financial decision-making. Accurate forecasts help stabilize economies amid uncertainty. Traditional models often struggle to capture nonlinear dynamics, high-dimensional data, and sudden regime shifts, leading to ineffective policy decisions. This study addresses these gaps by proposing HybridStack, a novel hybrid stacking ensemble combining gradient-boosted trees with regularized linear regression, optimized through Bayesian hyperparameter tuning and meta-learning strategies. The model is trained on a harmonized high-frequency panel dataset constructed by merging ten macro-financial sources covering U.S. inflation from 2006 to 2024. Empirical results show that HybridStack outperforms fifteen baseline models across seven metrics. On the held-out test set, it achieves a root mean squared error (RMSE) of 0.0006, a mean absolute error (MAE) of 0.0004, and a coefficient of determination (R2) of 0.9999. Notably, HybridStack achieves approximately a 99.92% reduction in RMSE relative to the best-performing prior U.S. study benchmark, a 94.7% reduction compared to the standalone Ridge baseline, and a 97.6% reduction relative to XGBoost on the held-out test set, establishing a new state-of-the-art benchmark within our dataset and evaluation framework. Further explainable AI (XAI) analyses using SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Partial Dependence Plots (PDP) highlight the primacy of short-term lags and rolling means as dominant predictors, with inflation expectations exhibiting episodic yet influential effects. This work contributes a transparent, empirically validated forecasting tool for policymakers and researchers, enhancing the reliability and interpretability of inflation predictions in volatile economic environments.