Hybrid ensembles for Robust UK inflation forecasting: An empirical evaluation with high-frequency covariates and structural-break adaptation

This research develops and empirically evaluates a hybrid ensemble framework for forecasting UK inflation, specifically, the CPI, CPIH and RPI indices, over the turbulent March 2022–March 2025 period. The research went further on traditional time-series methods (SARIMA, VAR), machine-learning algorithms (Random Forest, XGBoost), deep-learning architectures (LSTM, CNN+LSTM) and the forecasting library Prophet by constructing an ensemble that combines individual model forecasts via a linear meta-learner. Models are trained on monthly ONS data, with rolling‒window validation reserving the final 12 months for out-of-sample testing.

Individual SARIMA and VAR produced high forecast errors : SARIMA RMSE: CPI 3.94, R² –17.52; VAR RMSE: CPI 2.69, R² –7.66, confirming their limitations in the face of non-linearity issue and regime-shifting dynamics. Random Forest (CPI RMSE 2.37, R² –5.69) and XGBoost (CPI RMSE 2.05, R² –4.03) improved accuracy but yielded negative R² values, indicating poor generalization. LSTM further reduced error with CPI RMSE 2.05 and R² of 3.99, while the hybrid CNN+LSTM achieved positive fit for CPI RMSE 0.54 and R² 0.65; CPIH RMSE 0.72, R² 0.64; RPI RMSE 2.18, R² 0.47, demonstrating the benefit of combining convolutional feature extraction with sequence learning.

The main contribution is the ensemble of XGBoost, LSTM, Ridge Regression and Prophet, which delivers near-zero bias and high explanatory power across all indices: CPI RMSE 0.18 (R² 0.96), CPIH RMSE 0.16 (R² 0.98) and RPI RMSE 0.55 (R² 0.97). This represents over 90% error reduction compared to SARIMA and nearly 70% improvement over CNN+LSTM, demonstrating the value of model diversity and stacking in volatile macroeconomic environments. The ensemble’s robustness strongly capture short-term seasonality, long-term trends and non-linear interactions.