A Layered Decision-Aware Forecasting Architecture for Noisy Panel Environments

Machine learning forecasting systems in noisy, non-stationary, multi-entity environments are commonly optimized for pointwise predictive accuracy, although many downstream decisions depend on the relative ordering of entities rather than exact numerical forecasts. This creates a potential prediction–ranking–decision misalignment, where the model with the lowest forecast error may not generate the most useful ranking or decision outcome. This paper proposes a Layered Decision-Aware Forecasting Architecture (LDFA) that separates predictive calibration from decision-oriented transformation. Layers 1 and 2 generate heterogeneous forecasts through base learners, stacking, and adaptive weighting, while the main methodological contribution is Layer 3: a selective post-prediction controller based on source selection, activation gating, veto logic, and regime-aware scaling. LDFA is evaluated using daily data for 30 EGX firms from 2020 to 2025 under a leakage-controlled expanding-window walk-forward protocol. The results show that prediction, ranking, and decision metrics favor different models: the Ridge meta-model provides the strongest target-scale accuracy, the XGBRankerLTR learning-to-rank baseline achieves the strongest ranking information coefficient (IC), and the SoftmaxHybridGate achieves the highest Sharpe ratio. The Layer 3 controller slightly improves over the Ridge backbone in mean daily IC, Sharpe ratio, and maximum drawdown, but the gains are modest and not statistically conclusive. Ablation and sensitivity analyses show that uncontrolled correction is harmful, while selective activation is the main value-adding mechanism. Transaction-cost, dynamic-controller, bootstrap, and top-k robustness tests further indicate that LDFA is most useful under selective decision settings rather than as a universally dominant model. The findings position LDFA as a transparent decision-aware forecasting architecture for studying and managing prediction–ranking–decision trade-offs in noisy panel environments.