Capacity and Pricing Management with Demand Learning

Jian Chen, Zechao Li, Anyan Qi, Yining WangSSRN 2023 · DOI

TL;DR

This paper tackles the complex challenge of simultaneously managing product capacity and pricing in an environment with unknown demand. They introduce a novel "double-trisection" algorithm that achieves near-optimal performance by efficiently learning demand and adjusting capacity, outperforming simpler strategies. This work is significant because it provides a robust framework for real-world decision-making in dynamic markets where both supply and demand are uncertain.

What Problem Does This Paper Address?

Many businesses operate in environments where customer demand is not perfectly known. To succeed, they must make crucial decisions about how much capacity to invest in (e.g., manufacturing, inventory, fleet size) and what prices to set for their products. These decisions are intertwined: capacity affects the ability to meet demand, and pricing influences demand itself. Furthermore, capacity adjustments are often costly and irreversible, making them difficult to change once made. This paper addresses the problem of how a firm can effectively learn about demand on the fly and make optimal, integrated decisions about both capacity and pricing over time, especially when demand is uncertain.

Methods

The core of the paper's contribution lies in its novel algorithmic approach. They propose a "double-trisection" algorithm designed to handle the joint capacity and pricing problem under demand learning. This algorithm leverages a subroutine called "SAFETRISECT" which uses a trisection search to efficiently estimate optimal pricing strategies at a "safe" capacity level, minimizing the need for costly capacity adjustments during the learning phase. The main algorithm then uses this subroutine in a two-layer fashion: first, it searches for the optimal capacity level, and at each candidate capacity level, it uses SAFETRISECT to find the optimal price. This approach is theoretically proven to achieve an optimal regret rate, meaning its performance is as good as possible given the inherent uncertainty. The paper also extends this framework to scenarios with limited capacity adjustment opportunities and seasonal demand patterns.

Datasets

The authors evaluate their proposed algorithms using a testbed inspired by public operational and financial data from Yoplait, a yogurt brand. This testbed includes parameters for demand patterns, capacity-related costs, and initial capacity rates. The demand function is modeled using a logit form, and the noise term follows a uniform distribution.

Key Insights and Ideas

  • The surprising Ω(T³/⁵) lower bound: The paper establishes a lower bound on the regret that is surprisingly high, indicating that simple learning algorithms with O(√T) regret (common in simpler pricing problems) are not achievable here. This is attributed to the unique challenges of capacity management, particularly the cost and irreversibility of adjustments, which create delays between learning demand and acting on that knowledge.
  • Double-trisection for integrated decisions: The proposed "double-trisection" algorithm is a key innovation. It effectively combines price optimization with capacity adjustment by using a nested search strategy. This allows for efficient learning and decision-making in a complex, interdependent system.
  • "Safe" capacity levels for learning: The SAFETRISECT subroutine's use of a "safe" capacity level during price exploration is crucial. This strategy minimizes the impact of costly capacity adjustments during the learning phase, allowing for more accurate demand estimation before committing to significant capacity changes.
  • Limited adjustments don't cripple performance: The paper shows that even with a limited number of capacity adjustment opportunities, near-optimal performance can still be achieved if the number of opportunities is sufficiently large (e.g., K ≥ ln ln T). This suggests practical flexibility in applying the model.

Key Results

  • The proposed double-trisection algorithm achieves a cumulative regret of Õ(T³/⁵), which matches the established theoretical lower bound, indicating optimal asymptotic performance.
  • When capacity adjustments are limited to K opportunities, the regret upper bound is Õ(KT⁵⁻⁴⁻⁽ᴷ⁺¹/²⁾), which approaches the optimal rate as K increases.
  • For seasonal demand patterns, the algorithm achieves a regret upper bound of Õ(W²/⁵T³/⁵), where W is the number of seasons.
  • Numerical experiments demonstrate that the proposed algorithm significantly outperforms baseline "explore-then-commit" policies, especially for large time horizons.
  • The study also shows that practical choices for algorithm parameters (C₁ and C₂) can achieve satisfactory performance without needing to adhere strictly to conservative theoretical values.

Limitations and Future Work

  • Asymptotic analysis: The paper primarily relies on asymptotic analysis (large T), which might not fully capture the behavior for shorter time horizons. Developing exact analysis for multi-period problems remains a challenging direction.
  • Model assumptions: The model assumes certain properties of the demand and revenue functions (e.g., smoothness, concavity). Relaxing these assumptions could lead to more general applicability.
  • Exact analysis complexity: The paper highlights the significant complexity of exact analysis for multi-period joint capacity and pricing problems, even with a small number of decision variables. This complexity is a major hurdle for future research in this area.
  • Future directions: The authors suggest exploring the coordination of capacity and price adjustments with a limited number of price adjustment opportunities, as well as developing exact analysis for these problems.

Summary updated

BibTeX
@misc{chen2023capacity,
  author = {Jian Chen and Zechao Li and Anyan Qi and Yining Wang},
  title = {{Capacity and Pricing Management with Demand Learning}},
  year = {2023},
  doi = {10.2139/ssrn.4414361},
  howpublished = {SSRN},
}
Figure