AI-Based Forecasting Models for Critical Inventory: A Systematic Review

Abstract

Inventory management in industrial settings is shaped by high variability in consumption and replenishment lead times, which hinders planning. Given the dispersion of predictive approaches reported in the literature, a systematic review helps consolidate comparable evidence. Accordingly, this study conducts a systematic literature review (SLR) aimed at characterizing recurrent models and variables used for forecasting critical supplies, with a focus on the dairy industry and AI. The review followed Fink’s methodology; from 780 initial records, 39 eligible studies were selected. Results show that the most recurrent variables are concentrated in the temporal and operational components of demand, highlighting historical consumption and seasonality, complemented by lead time and, when traceability exists, available stock. In terms of approaches, neural networks/deep learning predominate especially recurrent architectures (LSTM/GRU/RNN) while traditional machine learning methods are used as comparative baselines. For evaluation, error metrics (MAE, RMSE, MSE, MAPE) prevail; however, for intermittent consumption a dual framework is proposed that separates occurrence (Macro-F1) and magnitude (MAE). These findings guide the design and evaluation of predictive models applicable to real-world scenarios with limited data availability