Hybrid Transformer-CatBoost Architectures with Boruta-SHAP Feature Selection and Monte-Carlo-Driven Cost-Benefit Optimization for Demand Forecasting in High-Variability Supply Chains: A Systematic Review (۲۰۲۰–۲۰۲۵)

Accurate demand forecasting in high-variability supply chains (e.g., cosmetics and fast-moving consumer goods) is critical for inventory optimization and bullwhip effect mitigation. The explosion of multi-source heterogeneous data has rendered traditional statistical models inadequate, driving the adoption of hybrid machine-learning architectures that combine attention-based Transformers with gradient-boosted decision trees such as CatBoost. Recent studies increasingly incorporate Boruta-SHAP pipelines for robust feature selection and interpretability, while Monte Carlo simulation coupled with genetic algorithms enables rigorous cost-benefit evaluation of forecasting policies. This paper presents the first systematic literature review of these integrated mechanisms published between ۲۰۲۰ and ۲۰۲۵. Thirty high-quality studies are classified into three main categories: hybrid predictive modeling, intelligent feature engineering, and economic validation frameworks. A detailed comparative analysis reveals significant gains in accuracy, interpretability, and net economic benefit, while highlighting remaining challenges in real-time deployment and generalization across volatile markets. Recommendations for future research are provided.

کلیدواژه ها:

Demand forecasting ، Supply chain management ، Transformer-Cat Boost ، Boruta feature selection ، SHAP explainability ، Monte Carlo