To create an AI-Driven Demand Forecasting Model for Perishable Goods in a Fragmented Retail Market. To develop an AI-Driven Demand Forecasting Model for Perishable Goods in a Fragmented Retail Market

Abstract

The operational problem of perishable goods is one of the most difficult demand forecasting problems because of the limited shelf life, the quick variation in demand, and the cost of having unsold stock go bad due to quality deterioration or the loss of sales and customer dissatisfaction due to stock outs. The situation is particularly difficult in fragmented retail markets, where small retailers, independent grocers, local distributors and informal supply chains commonly struggle to keep accurate records, lack effective digital systems, and have access to varying amounts of market information. In this paper, an artificial intelligence (AI) based demand forecasting framework is proposed which combines time-series learning, machine-learning features, hierarchical information sharing, and inventory-aware decision rules. The system is proposed to combine historical sales, price and promotion data, calendar effects, weather, local events, replenishment lead times, stock availability and the product shelf-life characteristics. It evaluates interpretable statistical baseline methods and machine learning techniques and a hybrid forecasting model, explicitly addressing censored demand due to stockout, data sparseness, new products, and store-level heterogeneity. The quality of forecasts is assessed by scaled error measures, probabilistic calibration and operational measures like waste, service level and lost sales. As no original retail data is provided, the paper does not include fabricated empirical results, but rather a research design and hypotheses. The framework states that the accuracy of the forecast is not an end in itself but should be considered within the framework of ordering decisions and perishability. A practical implementation ought to be simple to use, interpretable, respect privacy, and provide actionable forecasts even in the face of sparse data at the store level.