A Multi-Criteria Performance Evaluation of Contemporary Machine Learning Architectures for Stochastic Inventory Demand Characterization

DOI: https://doie.org/10.65985/JBSE.2025760527

Megha Thakar, Dr.Priya Swaminarayan

Keywords:

inventory demand forecasting, stochastic environments, supply chain management, traditional forecasting limitations, non-linear relationships, stock outs, machine learning architectures.

Abstract:

This study represents an effort to accurately forecast inventory demand in environments with stochasticity - a frequent issue in current supply chain management. Traditional forecasting methods do not recognize the inherently stochastic, non-linear, and often complex relationships between demand, inventory levels, and all of the associated conditions, which could lead to suboptimal inventory levels, the costs of holding inventory, and the possibility of stock outs. This study rigorously assessed a variety of ML architectures, including deep learning approaches (e.g., LSTMs, Transformers), and sophisticated ensembling methods (e.g., Gradient Boosting, Random Forest) that are designed for characterizing and forecasting stochastic inventory demand. Each architecture was assessed on a total of 10 operationally relevant metrics, examining performance in predicting accuracy (e.g., RMSE, MAE), computational efficiency, interpretation, and their ways to mitigate the effect of data noise and outliers, were included as part of deconstruction of operational value, and accuracy. Some deep learning models, for highly volatile demand, are better at capturing complex temporal dependencies but require specific operating environments that may not exist in practical applications. When considering all potential architectures, we identify trade-offs in performance and accuracy between the evaluated architectures, and found that the ensemble methods were attractive for combining both potential predictive power with computational tractability and beneficial attribution back to the attributes selected within the analysis of demand.

PDF