Aware-Hybrid Ultra Model: A Unified AI Framework for Multi-SKU Demand Forecasting and Intelligent Supply Chain Optimization

Abstract

Retail demand forecasting becomes increasingly unstable in multi-SKU, multi-store environments where behaviour is shaped by irregular promotions, heterogeneous store dynamics, and weak temporal persistence, causing conventional time-series models to degrade sharply. Addressing this challenge, the AWARE_HybridUltraModel is formulated as an adaptive, error-driven ensemble designed to operate effectively in settings that violate classical autocorrelation and seasonality assumptions. The framework integrates multiple heterogeneous regressors through an inverse-error weighting mechanism that assigns predictive influence according to each model’s empirical loss profile, while a CatBoost-based residual learner captures higher-order deviations left unresolved after ensemble fusion. This layered structure enables the system to remain stable across fragmented signals, nonlinear promotional effects, and SKU-specific variability. Extensive evaluation on a complete two-year multi-store dataset demonstrates that the model achieves MAE ≈ 1.96, RMSE ≈ 2.45, and R² ≈ 0.974, consistently outperforming Extra Trees, Gradient Boosting, XGBoost, Random Forests, CatBoost, Ridge, and MLP. In contrast, deep sequence architectures such as LSTM and TCN fail under the dataset’s weak sequential continuity, exhibiting catastrophic error amplification with RMSE values above 64. The proposed architecture produces a high-resolution, low-noise demand signal suitable for downstream operational tasks including replenishment planning, safety-stock calibration, and dynamic inventory optimisation. Overall, the AWARE_HybridUltraModel offers a rigorous, data-agnostic forecasting mechanism specifically tailored for modern retail streams where structural volatility and irregular behavioural drivers undermine the assumptions of conventional forecasting models.