This study proposes a multilevel stacking approach that extends Wolpert's (1992) stacked generalisation, aiming to minimise forecasting error, reduce systematic bias, and enforce temporal domain invariance. It incorporates walk-forward out-of-fold and step-based walk-forward validation, enriched metadata, and a domain adversarial neural network (DANN). The base learners XGB and LSTM function as residual and latent encoders for the Random Forest with DANN meta-learner. DANN is integrated in meta-learning to minimise task regression loss and maximize domain discrimination loss through a gradient reversal layer. This fusion improves domain transition stability, as the model’s high explanatory power with ^2 values exceed 0.98 across validation and test regimes. The MASE values below 0.07 on both regimes indicate strong predictive performance relative to a naïve benchmark, and NRMSE values around 0.025 confirm that the proposed stacking framework effectively captures temporal dependencies. Moreover, the model achieves a near-symmetric residual distribution, with a skewness of 0.37 on the validation set and -0.01 on the test set, suggesting that adversarial feature learning helps reduce asymmetric prediction errors and stabilise meta-learning. This proposed framework demonstrates empirically that stacking should not be evaluated solely on MAE/RMSE but on residual distribution behaviour under domain transition. Furthermore, SHAP analysis provides empirical evidence that enriched metadata significantly enhances the learning capacity, allowing the meta-learner to capture complex dependencies.