Developing an AI-Driven Demand Forecasting Model for Perishable Goods in Fragmented Retail Markets

Abstract

The problem of demand forecasting for perishables is an operational problem which is difficult because the product has a finite shelf life, the demand is subject to rapid variations and if the demand is not met, the stock is lost and dissatisfied customers. In decentralized retail markets, such as small shops, independent retailers, local wholesalers, and informal traders, these challenges are particularly acute, as they may lack proper records, digital systems, and market information. This paper presents an artificial intelligence (AI) based demand forecasting system comprising a time-series learning approach, machine-learning attributes, a hierarchical information sharing method, and inventory-based decision rules. Historical sales, price and promotion data, calendar effects, weather, local events, replenishment lead times, stock availability, and the characteristics of product shelf-life are integrated into the proposed system. A comparison of interpretable statistical baselines, machine-learning methods, and a hybrid forecasting model explicitly tackles censored demand due to stockout, limited data, new products, and store-level heterogeneity. Quality of the forecast is assessed using scaled error measures, probabilistic calibration and operational measures (waste, service level, lost sales). A research design and hypotheses are provided but no original retail data are provided. The concept of the framework is that the performance of a forecast should be evaluated with regard to ordering decisions and perishability, rather than simply as a statistical goal. A practical implementation should be simple to use, understandable, non-intrusive and provide practical forecasts even if information is scarce at specific stores.