- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper studies dynamic pricing when inventory is revealed before pricing, may depend on past observations, and censors demand. Under monotone expected sales, Threshold-UCB shares demand-threshold observations across inventory levels and achieves an expected regret bound of order T^(2/3), up to logarithmic factors. A posted-pricing reduction gives a matching lower bound. Experiments cover 48 synthetic environments.
Critical strengths
- The broader demand model is useful. The paper removes the linear-additive structure and inventory-support conditions of Xu et al. [X26], while establishing a matching upper and lower regret rate for its own model class. It correctly explains why the two models have different minimax rates.
- The threshold estimator has a clear interpretation: an observation is used only when inventory makes the corresponding demand indicator observable. Appendix B explains both the adaptive-sampling argument and why pooling avoids a separate estimation cost for each inventory bin. I also appreciate the separate tuning paths, 30 held-out paths, and variability reporting.
[X26] Xu, J., Wang, Y., Chen, X., and Wang, Y.-X. (2026). Dynamic Pricing with Adversarially-Censored Demands. arXiv:2502.06168, revised January 23, 2026. https://arxiv.org/abs/2502.06168.
Critical weaknesses
- The experiments do not use Algorithm 1 exactly. Appendix D.3 omits the zero-threshold term and changes the confidence widths. The omitted population revenue term is pΔS_p (0), which depends on price; this is not simply removing a constant shared by all actions. The modifications are disclosed, but their effect on the comparisons is unclear. A small comparison with Algorithm 1 as stated would help connect the numerical results to the theorem.
- Section 4 omits Double-Grid-UCB, so the experiments do not directly measure the benefit of sharing observations across inventory levels. Moreover, Appendix D.2 generates every inventory path before interaction; piecewise changes do not test inventories that react to past observations. These limits do not invalidate the theoretical result, but the empirical conclusions should distinguish what was tested from what the theorem permits.
Questions for authors
- On a small representative subset, how do Algorithm 1 as stated, its experimental variant, and Double-Grid-UCB compare under matched grid choices and tuning budgets? I am not asking for a repeat of all 48 environments.
- Is any history-dependent inventory experiment available? If not, please make clear that the numerical evidence concerns pre-generated inventory paths; the adaptive-inventory guarantee remains a theoretical result.
- A short discussion of two related settings would help. Bu et al. [B23] learn from offline censored data, whereas this paper receives online feedback. Ao et al. [A25] study a shared capacity budget over the horizon, rather than exogenous inventory each period. Please clarify these differences in information and resource constraints; a direct comparison of regret rates is not needed.
[B23] Bu, J., Simchi-Levi, D., and Wang, L. (2023). Offline Pricing and Demand Learning with Censored Data. Management Science, 69(2), 885–903. https://doi.org/10.1287/mnsc.2022.4382. [A25] Ao, R., Jiang, J., and Simchi-Levi, D. (2025). Learning to Price with Resource Constraints: From Full Information to Machine-Learned Prices. Advances in Neural Information Processing Systems, 38, 91243–91281. https://proceedings.neurips.cc/paper_files/paper/2025/hash/83995c2acef585ae0f0d647154cfdd85-Abstract-Conference.html.
2 comments
I will keep my score
Thanks for updating the paper.
We thank the reviewer for the time and constructive feedback. Below, we provide detailed responses to the reviewer’s comments. We have also submitted a revised PDF that includes a discussion and theoretical analysis of a new algorithm Shared Double-Grid-UCB (Algorithm 3), more related works, and additional experimental results for both Double-Grid-UCB and Shared Double-Grid-UCB. All changes are highlighted in red in the revised PDF.
Q1: How do Algorithm 1, the experimental implementation, and Double-Grid-UCB compare?
A1: Following your suggestion, we have included the results of the exact Algorithm 1, which keeps the zero threshold and uses the theoretical confidence radius log(2(K+1)JT2)/(2N), the experimental variant in Appendix D.3, which omits the zero threshold and uses the radius cconflog(2(K+1)JT/δ)/(2N) with δ=0.05, and Double-Grid-UCB under the same price grid. Specifically, we fixed K=J=22 and the confidence multiplier at one in the A, B, and C central-uniform additive settings. Each implementation uses its own stated confidence formula. Entries are mean final expected regret ± one sample standard deviation over the same 30 evaluation paths at T=10000. The results are shown in the following table. The two Threshold implementations perform similarly, and both have much lower regret than Double-Grid-UCB in all three settings.
Q2: Is there an experiment with history-dependent inventory?
A2: All reported experiments use inventory paths generated before the interaction, so they evaluate exogenous inventory sequences. We clarified this in Appendix D.2. Inventory that adapts to past sales is covered by our theory: Theorem 3.2 holds for any inventory sequence whose conditional law depends on the observed past, because the confidence bounds only require that eligibility for each threshold is determined before the current demand is realized (Lemma B.2).
Following your suggestion, we also ran experiments in which each seller’s sales determine its next inventory. We use the paper’s linear Demand A, Yt(p)=1.4−0.8p+εt, with independent shocks εt∼Unif[−0.1,0.1]. Each seller starts with low inventory γ1=L and observes γt before posting a price. If all available inventory sells in round t, the seller receives H units in round t+1. Otherwise, it receives L units. Thus, two algorithms facing the same demand shocks can generate different inventory paths through their pricing decisions.
For each inventory pair (L,H), we compared Threshold-UCB, fixed-horizon C20CB, Double-Grid-UCB, and Shared Double-Grid-UCB, a variant of Double-Grid UCB suggested by Reviewer 4Csy which uses each sale to update the revenue estimates for all lower inventory levels. Similar to all the experiments in the paper, we evaluated 20 configurations per algorithm by Bayesian search on five tuning paths and selected the one with the lowest mean final regret. We then evaluated each selected configuration on 30 separate paths of T=10000 rounds, using common demand shocks across algorithms. Each round’s benchmark is the best expected revenue at that algorithm’s realized inventory. The table reports mean final cumulative expected regret ± one sample standard deviation across the 30 paths.
The result shows that Threshold-UCB has the lowest mean final cumulative expected regret among the four algorithms in all four inventory settings.
Q3: How does the setting differ from Bu et al. and Ao et al.?
A3: Thank you for pointing out these related works. As the reviewer noted, Bu et al. (2023) learn from an offline data set of historical prices, inventory levels, and censored sales, whereas our learner chooses prices online and the inventory arrives sequentially and possibly adversarially. Ao et al. (2025) study dynamic pricing with a resource budget shared over the whole horizon, whereas in our model inventory is supplied before each period and the comparator optimizes revenue at each realized inventory, so there is no budget to allocate across periods. We have added this discussion to the related work in Section 1.1.