Abstract
The Newsvendor problem is a fundamental challenge in inventory management, requiring a balance between shortage and holding costs under uncertain demand. While Deep Neural Networks (DNNs) offer powerful predictive capabilities, they typically rely on symmetric quadratic loss functions that fail to capture the economic realities of asymmetric costs. This paper proposes a novel DNN architecture that directly optimizes the original Newsvendor loss function. By decomposing the objective into Rectified Linear Unit (ReLU) components, the model remains differentiable and trainable via standard optimization techniques. Empirical results on retail data and stress tests demonstrate that this approach provides superior generalization and significantly higher robustness to demand outliers compared to traditional L2-based models.
Introduction
In retail inventory management, the Newsvendor problem involves determining the optimal order quantity for perishable goods before a selling season. The decision-maker faces two competing costs: a shortage cost ($c_p$) for lost profit when demand exceeds supply, and a holding cost ($c_h$) for waste when supply exceeds demand. Traditionally, this is solved using historical demand distributions. However, modern demand is influenced by complex, non-linear features such as weather, holidays, and promotions.
While Deep Learning has been applied to demand forecasting, most implementations utilize a Quadratic (L2) loss function. This is often a matter of computational convenience, as L2 loss is smooth and easy to optimize. However, L2 loss is symmetric—treating over-ordering and under-ordering identically—and is highly sensitive to outliers. This paper addresses these shortcomings by integrating the original asymmetric Newsvendor loss function directly into the neural network training process.
Methodology: The ReLU Integration
The primary challenge in using the original Newsvendor loss function is its non-differentiability at the point where predicted demand equals actual demand. This typically prevents the use of back-propagation. The authors overcome this by demonstrating that the Newsvendor objective can be represented as a combination of two ReLU components.
As shown in Figure 1, the architecture feeds features $\mathbf{x}$ through a classic DNN to produce a demand prediction $f(\mathbf{x}, \mathbf{q})$. This prediction is then processed through two ReLU layers weighted by the cost parameters $c_p$ and $c_h$. Because ReLUs are differentiable almost everywhere and are standard in modern AI frameworks, the entire network can be optimized using the L-BFGS algorithm without requiring artificial "smoothing" of the loss function.
Experimental Results and Robustness
The model was evaluated using real-world retail data (Foodmart 1999) and simulated stress tests. A key finding is the model's performance under "Simulation 2," where demand outliers were artificially introduced to test robustness.
As illustrated in Figure 4, the quadratic loss function "chases" outliers, leading to a dramatic spike in test error as the model over-adjusts to extreme values. In contrast, the proposed model using the original loss function remains stable. By maintaining an L1-style approach, the network ignores the "noise" of outliers and focuses on the underlying demand patterns, leading to much better generalization in practical business scenarios.
Conclusion
Integrating the original Newsvendor loss function into Deep Learning models provides a more economically grounded approach to inventory management. The "ReLU trick" allows for efficient optimization while preserving the essential asymmetry of the problem. The resulting models are not only more accurate across various cost ratios but also significantly more robust to the demand volatility common in real-world retail environments. Future research will focus on extending this framework to multi-product scenarios involving substitution and complementary effects.
Summary updated