Explainable Agentic Pipelines for Accurate Demand Forecasting

In today’s competitive retail environment, accurate demand forecasting is vital for minimizing costs, enhancing service levels, and optimizing inventory management. Traditional forecasting methods frequently underperform in highly dynamic settings, particularly when both predictive accuracy and interpretability are necessary. This thesis introduces a hybrid framework that combines advanced machine learning models with eXplainable Artificial Intelligence (XAI) methods within an agentic AI architecture. Random Forest (RF) and eXtreme Gradient Boosting (XGBoost), which are ensemble-based predictive models used for classification and regression, serve as the primary predictive models. SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) deliver interpretability by explaining how features contribute to individual predictions. This enables managers to comprehend model behavior and foster trust in AI-driven forecasts. This work presents the AI Refinery, a modular system that orchestrates forecasting, evaluation, and refinement through iterative processes. Forecasts are generated and benchmarked against reference baselines, while explainability techniques assess feature contributions and guide improvements. Additional modules support data analysis, similarity matching, and the integration of external information, ensuring adaptability and model diversity. The results demonstrate that integrating predictive performance with explainability bridges the gap between advanced analytics and managerial decision-making. By embedding governance and trust into the forecasting pipeline, this work contributes a scalable and extensible approach to demand forecasting, improving both technical robustness and practical utility in real-world supply chain contexts.