[Paper Review] Stock-out Prediction in Multi-echelon Networks

Afshin Oroojlooyjadid, Lawrence Snyder+1 more|arXiv (Cornell University)|Sep 20, 2017

Supply Chain and Inventory Management9 references3 citations

TL;DR

This paper proposes a deep neural network (DNN)-based approach to predict stock-outs in multi-echelon supply chains using historical inventory and demand data, without requiring knowledge of demand distributions. The method achieves up to 99% prediction accuracy, significantly outperforming naive benchmarks, especially in systems with dependent demand and multi-period forecasting.

ABSTRACT

In multi-echelon inventory systems the performance of a given node is affected by events that occur at many other nodes and in many other time periods. For example, a supply disruption upstream will have an effect on downstream, customer-facing nodes several periods later as the disruption "cascades" through the system. There is very little research on stock-out prediction in single-echelon systems and (to the best of our knowledge) none on multi-echelon systems. However, in real the world, it is clear that there is significant interest in techniques for this sort of stock-out prediction. Therefore, our research aims to fill this gap by using deep neural networks (DNN) to predict stock-outs in multi-echelon supply chains.

Motivation & Objective

• To address the lack of research on stock-out prediction in multi-echelon supply chain networks, where system-wide disruptions cascade through nodes.
• To develop a data-driven method that predicts near-future stock-outs using only historical inventory and demand data, without assuming known demand distributions.
• To evaluate the performance of deep learning against naive baseline methods in predicting stock-outs across various network topologies and demand correlation structures.
• To assess the model’s ability to predict stock-outs multiple periods in advance and under varying levels of demand dependence.
• To explore the impact of incorporating demand dependence information into the DNN architecture for improved prediction accuracy.

Proposed method

• A deep neural network (DNN) is trained on historical data of inventory levels, in-transit quantities, and demand to predict whether a node will face a stock-out in the near future.
• The model is evaluated against three naive baseline methods: random prediction, constant prediction, and prediction based on current inventory only.
• For multi-period prediction, the DNN is modified to output q predictions for the next q time periods, with separate models tested for each horizon length.
• A weighted DNN variant is introduced to account for demand dependence, where correlation between items is modeled as a learned weight in the input layer.
• The model is trained and tested on synthetic data generated from real-world network topologies, including serial, OWMR, and distribution networks with varying demand correlation structures.
• Performance is evaluated using false-positive and false-negative rates, with accuracy computed across multiple test instances and prediction horizons.

Experimental results

Research questions

• RQ1 Can a deep learning model predict stock-outs in multi-echelon supply chains more accurately than naive baseline methods when no demand distribution is known?
• RQ2 How does incorporating demand dependence (positive or negative correlation between items) affect the prediction accuracy of the DNN model?
• RQ3 What is the predictive accuracy of the DNN model when forecasting stock-outs multiple periods in advance, and how does it degrade over longer horizons?
• RQ4 Does the weighted DNN variant, which accounts for demand dependence, significantly improve prediction performance compared to the standard DNN?
• RQ5 How does the model perform across different network topologies, including complex non-tree structures?

Key findings

• The DNN model achieves an average prediction accuracy of 99% in serial networks, significantly outperforming the naive baseline methods.
• In distribution networks with multi-item dependent demand, the weighted DNN achieves 99% accuracy, compared to 95% for the independent-demand case, demonstrating the value of modeling demand dependence.
• For complex, non-tree network topologies, the DNN maintains an average accuracy of at least 91%, indicating robustness across diverse supply chain structures.
• Multi-period prediction accuracy decreases with longer horizons: predictions for days 4–7 are below 90% when forecasting 7 days ahead, and earlier predictions degrade when using longer horizons.
• Running separate DNN models for each prediction period (e.g., 2-day, 3-day, 7-day horizons) yields higher accuracy than a single multi-output model, especially for earlier time steps.
• The model performs well even when predicting stock-outs below a threshold other than zero, indicating flexibility in defining stock-out conditions.