Stochastic approach for price optimization problems with decision-dependent uncertainty

Yuya Hikima, Akiko Takeda

TL;DR

This paper tackles the complex problem of optimizing prices when demand uncertainty isn't fixed but actually depends on the price itself. They introduce a new, general mathematical framework and a novel algorithm that can handle this decision-dependent uncertainty, outperforming existing methods in experiments. This is crucial for real-world applications where pricing strategies directly influence how predictable customer behavior becomes.

What Problem Does This Paper Address?

In many business scenarios, setting the right price is key to maximizing revenue. However, demand for products is rarely perfectly predictable; it's usually a "stochastic" (random) process. Traditional approaches often assume this demand uncertainty is independent of the price set. For example, the demand might be a base amount plus some random noise, and this noise doesn't change whether the price is high or low.

This paper argues that this assumption is often unrealistic. In reality, demand uncertainty can depend on the price. Think about a niche product: if the price is set very high, demand might be low but also highly unpredictable. Conversely, a price close to competitors might lead to more predictable demand. Existing methods that try to handle this "decision-dependent uncertainty" often make very specific assumptions about the demand distribution or the problem structure, limiting their applicability. This paper aims to provide a more general framework and a robust algorithm to address this challenge across a wider range of pricing problems.

Methods

The core of the paper lies in two main contributions:

  1. A General Non-Convex Stochastic Optimization Formulation: They propose a mathematical model that captures revenue maximization where the probability distribution of demand is a function of the price. This formulation is designed to be general, allowing for non-convexity (meaning the "best" solution isn't always found by simple gradient ascent) and decision-dependent uncertainty.
  2. Projected Stochastic Gradient Descent with Variance Reduction: To solve this challenging optimization problem, they develop a novel algorithm. It's a projected stochastic gradient descent method, which is a standard technique for optimization. However, they introduce a key innovation: an unbiased stochastic gradient estimator that incorporates a "variance reduction parameter." This parameter, inspired by techniques in reinforcement learning, helps to make the gradient estimates more stable and reliable, leading to faster convergence and better solutions, especially in high-dimensional problems. They also propose efficient ways to compute these gradients, particularly for multi-agent scenarios.

Datasets

The authors evaluated their proposed method using:

  • Synthetic Experiments: They generated 20 random problem instances to test their algorithm under various conditions.
  • Real Retail Data: They used sales data from a supermarket service provider in Japan, specifically focusing on top-selling confectionery products over different weeks.

Key Insights and Ideas

  • Price-Dependent Uncertainty is Real and Important: The paper strongly emphasizes that assuming demand uncertainty is independent of price is a significant simplification that can lead to suboptimal pricing strategies.
  • Generality is Key: Their proposed formulation is designed to be broadly applicable, avoiding the restrictive assumptions of many prior works.
  • Variance Reduction is Crucial for Stability: The introduction of a variance reduction parameter in the stochastic gradient is a critical component for achieving reliable convergence in non-convex, decision-dependent settings.
  • Gradient-Based Methods Can Work for Complex Problems: By carefully designing their stochastic gradient estimator, they enable gradient-based optimization to be effective even when the underlying problem is non-convex and the demand distribution is complex.

Key Results

  • Outperforms Baselines: In both synthetic experiments and simulations with real retail data, their proposed method consistently achieved higher total revenues (measured as Negative Expected Revenue, NER) compared to several established baseline methods, including L2-Regularized Repeated Gradient Descent, Bayesian Optimization, SPSA, and PSD-AD.
  • Improved Convergence: The variance reduction technique in their stochastic gradient estimator leads to more stable convergence, avoiding the "unreasonably high prices" or "low-quality local solutions" observed in some baseline methods.
  • Theoretical Guarantees: The proposed algorithm is theoretically guaranteed to converge to an ε-stationary point, providing a strong foundation for its practical application.

Limitations and Future Work

  • Stationary Point vs. Global Optimum: The current method guarantees convergence to a stationary point, not necessarily the globally optimal solution. The authors suggest future research could explore methods for finding global optima, perhaps by incorporating multi-start techniques or building faster Bayesian optimization approaches under more specific assumptions.
  • Assumption Relaxation: The paper relies on several assumptions (e.g., differentiability of the probability density function). Future work could investigate how the method performs when these assumptions are relaxed, potentially by using smoothing techniques.
  • Scalability for Very High Dimensions: While the method is designed for high-dimensional problems, the computational cost of calculating minibatch gradients can still be a bottleneck. Specialized approaches for multi-agent applications are proposed to mitigate this, but further research into computational efficiency is always valuable.

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},
}

Loading related content...

Figure