Drawing on the ensemble principle validated by Huber and Stuckenschmidt~\cite{ref12}, the hybrid forecast is: \[ \hat{y}_t = \alpha \cdot \text{XGBoost}(X_t) + (1-\alpha) \cdot \text{LSTM}(X_{t-k:t}) \] where $\hat{y}_t$ is predicted demand at period~$t$, $X_t$ the contemporaneous feature vector, and $X_{t-k:t}$ the historical sequence over $k=12$ preceding periods. The blending coefficient $\alpha=0.4$ was selected by grid search on the validation fold.

Based on the ensemble approach established by Huber and Stuckenschmidt, we construct our hybrid forecast using the following formula: \[ \hat{y}_t = \alpha \cdot \text{XGBoost}(X_t) + (1-\alpha) \cdot \text{LSTM}(X_{t-k:t}) \] In this equation, $\hat{y}_t$ represents the forecasted demand for time period $t$, while $X_t$ denotes the feature vector for the current period. The term $X_{t-k:t}$ captures the historical data sequence spanning $k=12$ previous time periods. We determined the optimal blending parameter $\alpha=0.4$ through systematic grid search evaluation using our validation dataset.