Algorithmic Forecasting of Fresh Demand: A Comparative Analysis of ML Approaches, Selection of Metrics (WMAPE/Pinball), and Resilience to Shocks

Abstract

The rising instability of fresh food markets, which has been marked by demand shocks, climatic uncertainty, promotion-driven variability, and logistics volatility, has hugely increased the economic impacts of prediction errors in perishable supply chain. In Fresh categories, demand misalignment can be converted into instantaneous waste production, stock-outs, loss of margins and inefficient capital usage. In spite of these structural limitations, a significant portion of retail forecasting systems to this day rest on linear statistical models and heuristic safety buffers which do not work very well under nonlinear demand dynamics and shock regimes. This paper focuses on algorithmic demand forecasting as a structural format of coordination in the digital Fresh ecosystems instead of a strictly technical-analytic instrument. The study performs a comparative analysis of the classical time-series (SARIMA) models with the machine learning (Gradient Boosting XGBoost, and Recurrent Neural Networks LSTM) on the longitudinal retail data (120 stores, 24 months). The analytical structure incorporates weighted mean forecast accuracy (WMAPE), probabilistic calibration (Pinball Loss), shock sensitivity test, and simulation of financial impact.