Regret in the Newsvendor Model with Demand and Yield Randomness

Zhi Chen, Weijun Xie

Abstract

Traditional inventory management models often rely on the assumption that the underlying probability distributions of demand are known. However, in real-world supply chains, decision-makers face both distributional ambiguity and supply-side uncertainty. This paper presents a data-driven framework for the classic newsvendor problem that incorporates both demand and yield randomness. By leveraging distributionally robust optimization (DRO) with a type-$\infty$ Wasserstein distance, we develop a minimax regret model that provides high-performance ordering decisions even with limited historical data.

Introduction

The "newsvendor problem" is a fundamental challenge in operations research: determining the optimal order quantity for a perishable product under uncertain demand. While classical models assume a known probability distribution, modern supply chains are characterized by "distributional ambiguity," where only a limited set of historical data points is available.

Furthermore, most existing literature focuses solely on demand uncertainty, ignoring "yield randomness"—the risk that the quantity delivered by a supplier deviates from the quantity ordered due to manufacturing defects, logistics disruptions, or environmental factors. This research addresses the complexity of making optimal inventory decisions when both demand and supply are stochastic and their joint distribution is unknown, relying instead on an empirical, data-driven approach.

Methodology

Distributionally Robust Optimization (DRO)

To address distributional ambiguity, the authors employ Distributionally Robust Optimization. Rather than optimizing for a single estimated distribution, DRO seeks a solution that performs well against the "worst-case" distribution within a defined ambiguity set—a collection of all plausible distributions consistent with the observed data.

Type-$\infty$ Wasserstein Distance

The ambiguity set is constructed using the type-$\infty$ Wasserstein distance. This metric creates a robust "ball" around the empirical data points. A distribution is considered a candidate for the true distribution if it lies within this ball. This specific metric is chosen for its superior performance in scenarios with small sample sizes, providing a more stable boundary than traditional moment-based methods.

Minimax Regret and the Hurwicz Criterion

While many robust models focus on maximizing the worst-case profit (maximin), this approach can be overly pessimistic. Instead, this paper utilizes the Minimax Regret criterion. Regret is defined as the difference between the actual profit realized and the profit that could have been achieved with perfect hindsight. By minimizing the maximum possible regret, the model ensures a balanced performance across various market realizations.

Additionally, the framework is extended to include the Hurwicz Criterion, allowing decision-makers to calibrate their level of conservatism. By adjusting a weight factor ($\lambda$), the model can balance optimism (maximax) and pessimism (maximin) to suit specific risk appetites.

Theoretical Insights

The Complexity of Yield Randomness

The inclusion of random yield significantly increases mathematical complexity. In demand-only models, regret functions typically exhibit simple piecewise concavity or convexity. However, yield randomness introduces non-linearities that require a sophisticated "splitting" technique to maintain tractability.

The "Equal Regret" Property

A significant theoretical contribution of this work is the identification of the "Equal Regret" property. The authors prove that the optimal minimax regret order quantity occurs at the point where the maximum regret of under-ordering (lost opportunity) exactly equals the maximum regret of over-ordering (excess cost).

Computational Efficiency

Despite the underlying complexity of Wasserstein DRO, the authors demonstrate that the minimax regret objective is a one-dimensional convex function. This allows the optimal order quantity to be identified using a Golden Section Search algorithm. In practice, the computational requirement is negligible, with solutions found in less than 0.1 seconds.

Experimental Evaluation

The model was evaluated using synthetic datasets designed to simulate diverse supply chain environments:

  • Joint Distributions: The FGM Copula was used to model varying levels of correlation (negative, uncorrelated, and positive) between demand and yield.
  • Distributional Variety: Testing included scenarios where the true underlying demand followed Uniform, Exponential, and Normal distributions.
  • Data Scarcity: To test robustness, the model was trained on very small sample sizes ($N=5, 15, 50$).
  • Performance Metric: Out-of-sample performance was validated against $10^7$ random testing samples.

Key Results

  1. Superiority of Minimax Regret: In numerical experiments, the minimax regret model outperformed alternative criteria (maximin, maximax, and Hurwicz) in 33 out of 36 tested scenarios.
  2. Robustness to Small Samples: The type-$\infty$ Wasserstein approach proved significantly less sensitive to small training sets than traditional methods, which often suffer from bias when initial mean or variance estimates are inaccurate.
  3. Convergence: As the number of historical samples increases, the data-driven order quantity converges to the theoretical optimum, proving the model's consistency.
  4. Uncertainty Reduction: The ratio of minimax regret to the total possible profit range was consistently low, indicating that the model effectively narrows the window of potential loss.

Conclusion and Future Work

This research provides a robust, computationally efficient framework for the newsvendor problem under dual uncertainty. By focusing on regret rather than absolute worst-case outcomes, the model offers a more practical tool for inventory managers facing ambiguous market conditions.

Future research directions include extending this framework to multi-period inventory systems with carry-over stock, exploring multi-item portfolios with correlated demands, and integrating joint inventory and pricing decisions into the robust optimization framework.

Summary updated

BibTeX
@inproceedings{chen2021regret,
  author = {Zhi Chen and Weijun Xie},
  title = {{Regret in the Newsvendor Model with Demand and Yield Randomness}},
  booktitle = {Production and Operations Management},
  year = {2021},
  url = {https://openreview.net/forum?id=5cHPaaYanE},
}
Figure