acceptodds
Under review as a conference paper at ICLR 2027

Learning While Pricing Fairly: Queue-Regularized Exploration for Pricing Experiments

Abstract

Pricing experiments generate information about demand, but the prices used for learning are also the offers that different customer groups receive. We study how a seller can reduce differences between group-average prices while retaining the random variation needed for demand estimation. We extend Random Shock Design (RSD) with a fairness penalty scaled by virtual queues. This fairness-regularized RSD (FRSD) uses past disparities to correct the next price, then adds an independent shock for demand estimation. Under bounded demand and a well-conditioned regression design, we bound terminal price disparity and establish vanishing disparity at zero tolerance with diminishing shocks, while demand estimation remains consistent. The revenue analysis separates learning error from price adjustments; a fair benchmark and separate lower bounds identify the cost of terminal fairness. Experiments compare FRSD with RSD, fixed penalties, and current-gap penalties. Synthetic results include a 92% reduction in mean disparity alongside a 2.4% increase in mean revenue relative to RSD with the same estimation and exploration settings. In the Online Retail-calibrated simulation, FRSD reduces disparity by 19% with small costs in revenue and estimation accuracy.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
fairness, pricing experiments, online learning, exploration, virtual queues

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jy7UOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how a seller can reduce differences between group-average prices while retaining the random variation needed for demand estimation. The authors propose an approach called FRSD to extend Random Shock Design (RSD) with the introduction of a fairness penalty scaled by virtual queues.

Critical strengths

The main point of strength of this work is the terminal-disparity guarantee while preserving consistent demand estimation. The mathematics seems solid under the declared assumptions and the experimental part introduces useful ablations and evaluations.

Critical weaknesses

The experimental comparison is limited to RSD and to two ablations of the penalty (fixed and current-gap). There is no comparison with existing fair dynamic pricing methods, such as Cohen et al. (2025) or Doubly Fair Dynamic Pricing (Xu et al., 2023), nor with primal-dual controllers from the constrained bandit literature that the paper itself cites (Guo and Liu, 2024; Bernasconi et al., 2025).

The improvement over RSD on the Retail experiment is relative to the RSD with matched estimation and exploration parameters. Other RSD configurations achieve higher revenue and lower disparity, although with higher estimation error. Comparisons under common fairness and estimation budgets would help to better assess FRSD's advantage.

Demand coefficients are fixed, and changing demand is not evaluated or explicitly discussed. This is particularly relevant because both estimation and fairness correction retain historical information, potentially slowing adaptation after a regime change.

Considering the related work, I suggest to include Chen, Lyu, Zhang and Zhou, "Fairness-aware Online Price Discrimination with Nonparametric Demand Models". Moreover, the distinctions from the already cited works of Cohen et al. (2025) and Xu et al. (2023) would also benefit from a more technical discussion.

Questions for authors

Could you provide a technical justification for excluding a test in a common experimental setting with the existing similar approaches highlighted in the related work?

Does FRSD admit an upper bound on its excess loss relative to the optimal fair benchmark? If not, could you clarify what conclusions about the optimality of the policy can be drawn from the regret lower bound?

Reviewer_hqLzOpenReview ↗
Rating
3
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies pricing experiments that seek to reduce differences in average prices offered to different customer groups while learning demand. It extends Random Shock Design (RSD) with a fairness penalty that increases as past price disparities accumulate. The authors establish demand-estimation consistency and vanishing terminal disparity at zero tolerance, analyze revenue loss. Numerical study includes synthetic and retail-calibrated simulations.

Critical strengths
  • The method is simple to implement. Although it is a simple modification of existing method RSD, the paper explains clearly the new algorithmic component and points out the connection. The terminal-fairness result directly connects the proposed pricing rule to declining group-price disparities while preserving demand learning.

  • The experiments include comparisons with simpler penalties, separate training and test runs, and longer-horizon evaluations. The authors also report settings in which the proposed method performs less favorably.

Critical weaknesses
  • The demand-learning method closely follows existing RSD. Although the terminal-fairness guarantee is useful, the paper does not establish whether FRSD achieves fairness at a revenue cost close to that of the best fair policy. Theorem 2 leaves the cost of price adjustments uncontrolled. It is therefore unclear how much of its revenue loss is necessary for fairness and how much results from the proposed correction.

  • The experiments give mixed evidence or the benefit of accumulating past disparities. In Retail, the reported improvement is relative to RSD with matching exploration and regression settings. Compared with training-selected RSD, FRSD has higher disparity and lower revenue, although better estimation accuracy.

Questions for authors

Can you quantify FRSD’s revenue loss relative to the best fair policy, and clarify whether the excess loss per arrival vanishes as the horizon increases?

Reviewer_3jXpOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the problem of reducing differences in group-average prices while maintaining informative pricing experiments. The paper proposes fairness-regularized Random Shock Design, which uses accumulated disparities to adjust prices before adding independent exploration shocks. It establishes demand-estimation and terminal-fairness guarantees under stated assumptions and evaluates the method in synthetic and retail-calibrated simulations

Critical strengths
  1. Applying the fairness correction before drawing the exploration shock is a well-motivated design choice. It preserves the random variation used to identify demand while allowing historical disparities to influence prices

  2. The terminal-fairness result goes beyond simply assuming that the virtual queues remain controlled. The proof connects the implemented pricing rule to declining disparity while allowing demand estimation to remain consistent

  3. The experiments use matched randomness, separate parameter-selection runs, and comparisons with simpler penalties. The paper also reports settings where its method performs worse, which gives a more balanced picture of the trade-offs

Critical weaknesses
  1. The regret analysis leaves the cumulative cost of price adjustments as a separate quantity. It therefore does not establish that FRSD approaches the revenue of the optimal fair benchmark, even when its disparity vanishes

2.The benefit of accumulating disparities is not consistent across settings. Simpler penalties achieve smaller disparities in Synthetic C2 under the primary selection rule, while the Retail comparison with the current-gap penalty is statistically unresolved

  1. The Retail experiment uses transaction data to calibrate a simulator with an imposed price slope. Together with the common-slope assumption and aggregate fairness criterion, this limits what the results establish about heterogeneous customer responses or customer welfare
Questions for authors

NA

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.