Forecasting Under Supply Chain Volatility: Evaluating Machine Learning as a Pathway to Agentic Demand Planning

Abstract

Demand forecasting forms a critical component of supply chain planning. Forecasting errors directly impact inventory levels, replenishment decisions, service rates, and operational costs. Conventional statistical approaches such as autoregressive and exponential smoothing models remain effective for stable demand. However, their accuracy deteriorates sharply under non-linear, highly volatile, and rapidly changing market conditions. Recent advances in machine learning (ML) ensembles and Agentic Artificial Intelligence (AI) offer new pathways. These pathways support the development of highly adaptive demand planning systems. This paper investigates the performance of advanced ML architectures against traditional statistical baselines. It examines performance under varying degrees of supply chain volatility. We employ a PRISMA-based systematic literature review. The review identifies 89 relevant enterprise supply chain studies. From these studies we extract two high-fidelity enterprise cases for detailed comparative analysis. The first case evaluates forecasting architectures in a complex FMCG wholesale environment. An integrated DeepAR-XGBoost ensemble with Bayesian hyperparameter optimization reduces the Weighted Absolute Percentage Error (WAPE). The ensemble lowers WAPE from 0.40 for exponential smoothing to 0.17. This change produces a 57.5 percent reduction. The second case examines industrial demand volatility profiles. ML-based forecasting achieves improvements that range from 7.49 percent in stable environments to 83.40 percent over ARIMA under extreme volatility. These empirical results indicate that the operational value of ML-based forecasting depends heavily on demand characteristics. Exponential improvements occur under non-stationary conditions. Building on these findings, the study proposes a four-dimensional framework. The framework comprises Perception, Learning, Autonomy, and Execution. It conceptualizes how highly accurate ML pipelines serve as the foundational perception layer. This layer is required to transition toward Agentic Demand Planning. Ultimately, this research provides robust empirical evidence on volatility-dependent forecasting. It also supplies a strategic roadmap for integrating predictive intelligence into autonomous supply chain decision-making.