Stochastic approach for price optimization problems with decision-dependent uncertainty

Yuya Hikima, Akiko Takeda

Abstract

Price determination is a cornerstone of revenue management, yet its effectiveness often hinges on how well models account for the stochastic nature of demand. Traditional pricing models frequently assume that demand uncertainty is independent of the price set by the decision-maker. However, in many real-world scenarios, price directly influences the volatility of demand. This paper addresses price optimization under decision-dependent uncertainty (DDU) by proposing a general non-convex stochastic optimization framework. We derive an unbiased stochastic gradient estimator that incorporates a variance reduction parameter to ensure stable convergence. Our approach, evaluated on both synthetic and real-world retail data, demonstrates superior revenue generation and scalability compared to existing baselines.

Introduction

In revenue management, the standard model for demand is often expressed as $Demand = d(p) + \xi$, where $d(p)$ is a deterministic function of price and $\xi$ is a random noise variable. A critical, yet often unrealistic, assumption in these models is that $\xi$ is independent of the price $p$.

In practice, demand uncertainty is rarely static. For instance, setting a price close to a competitor’s may lead to high demand volatility as consumers fluctuate between options, whereas a significantly higher or lower price might result in more predictable (albeit different) demand levels. This phenomenon is known as decision-dependent uncertainty. While some specialized models have addressed this, they often rely on application-specific assumptions or require the objective function to be strongly convex—a property rarely found in complex, multi-product retail environments. This research introduces a scalable, general-purpose algorithm designed to find stationary points in non-convex pricing problems where the demand distribution shifts in response to price changes.

Methodology

1. The Stochastic Gradient Framework

The researchers employ a Projected Stochastic Gradient Descent (PSGD) approach. The primary mathematical hurdle in DDU is that when the probability distribution of demand $D(x)$ depends on the decision variable $x$, the gradient of the expected revenue is not simply the expectation of the revenue's gradient. To solve this, the authors derive a novel, unbiased stochastic gradient estimator using the "score function" method (the gradient of the log-likelihood of the demand).

2. Variance Reduction via the "Baseline Trick"

Drawing inspiration from the REINFORCE algorithm in Reinforcement Learning, the authors introduce a variance reduction parameter, $\delta$. The estimator is formulated as:

$$g(x, \xi, \delta) = \nabla_x f(x, \xi) + (f(x, \xi) - \delta) \frac{\nabla_x \text{Pr}(\xi|x)}{\text{Pr}(\xi|x)}$$

By dynamically updating $\delta$ to stay close to the expected revenue using an Online Gradient Descent (OGD) sub-algorithm, the "noise" in the gradient estimate is significantly reduced. This stabilization allows the algorithm to converge much more efficiently than standard stochastic methods.

3. Scalability for Multi-Agent Systems

For retail applications involving numerous products and buyers, calculating gradients can become computationally intensive. The paper provides a specialized version of the algorithm that exploits the mathematical structure of multinomial and binomial distributions, making it feasible to optimize high-dimensional price vectors in real-time.

Experimental Evaluation

Datasets

The proposed method was tested against two distinct environments:

  • Synthetic Data: Controlled environments with varying numbers of products ($n=10$ to $40$) and buyers ($m=100$ to $400$) to evaluate scalability.
  • Real-World Retail Data: Sales data for 50 confectionery products provided by KSP-SP Co., Ltd., a Japanese supermarket service provider. This dataset captured actual price fluctuations and sales rankings over the course of 2022.

Performance Metrics

The researchers compared their method against several industry standards:

  • Bayesian Optimization (BO)
  • Simultaneous Perturbation Stochastic Approximation (SPSA)
  • L2-Regularized Repeated Gradient Descent (L2-RGD)

Key Insights and Results

  • Superior Revenue Generation: Across both synthetic and real-world trials, the proposed method consistently achieved the lowest Negative Expected Revenue (NER), translating to the highest total revenue among all tested algorithms.
  • Computational Efficiency: While Bayesian Optimization is effective for low-dimensional problems, its computational cost becomes prohibitive as the number of products increases. The proposed stochastic gradient method maintained high performance even as the problem scaled.
  • The Necessity of Variance Reduction: Experiments showed that versions of the algorithm without the dynamic $\delta$ parameter (or with $\delta$ set to zero) suffered from poor stability and slower convergence, validating the importance of the RL-inspired baseline trick.
  • Handling Non-Convexity: Unlike previous "retraining" methods that require strong convexity, this framework successfully identifies stationary points in non-convex landscapes, which are typical in multi-product retail scenarios.

Limitations and Future Work

While the method is robust, it is designed to find a stationary point (a local optimum). In highly complex non-convex landscapes, this may not be the global optimum. Future iterations may incorporate "multi-start" techniques to explore the search space more broadly. Additionally, the current model requires the probability density function (PDF) of demand to be differentiable. Future research will focus on relaxing this requirement to accommodate "jumpy" or discrete demand changes and scenarios where the PDF is not perfectly known.

Conclusion

By bridging the gap between Reinforcement Learning techniques and revenue management, this paper provides a powerful new tool for price optimization. The ability to account for decision-dependent uncertainty without sacrificing scalability or requiring strict convexity marks a significant advancement for both theoretical research and practical retail applications.

Summary updated

Abstract

Highlights•Formulation of general pricing problems by decision-dependent uncertainty.•The formulation can capture the price-dependent uncertainty of demand.•Projected stochastic gradient descent with low-variance stochastic gradients.•Effective computation of minibatch gradients in multi-agent applications.
BibTeX
@article{hikima2025stochastic,
  author = {Yuya Hikima and Akiko Takeda},
  title = {{Stochastic approach for price optimization problems with decision-dependent uncertainty}},
  journal = {Eur. J. Oper. Res.},
  year = {2025},
  doi = {10.1016/j.ejor.2024.12.023},
}
Figure