acceptodds
Under review as a conference paper at ICLR 2027

Safe Contextual Pricing under Censored Feedback: Joint Inventory and Allocation with Online Conformal Guarantees

Abstract

Many online decision problems are simultaneously profit-seeking and safety-constrained: an agent must minimise regret while keeping a downstream operational risk below a prescribed level, using only censored feedback. We study this tension in online contextual pricing with multi-supplier inventory and price-dependent allocation, where the operational risk is the stockout rate. The seller observes only whether each customer purchases, must commit inventory across heterogeneous suppliers before demand realises, and faces an unknown and possibly non-Lipschitz demand-noise distribution. Our algorithm CP-CPA prices by maximising an optimistic profit on a residual grid, provisions the estimated profit-optimal inventory, and raises it to an online conformally calibrated demand quantile whenever the coverage target requires more. Its regret against the unconstrained clairvoyant is , the optimal rate under agnostic noise, plus exactly the holding cost of the safety inventory added by the conformal layer; this cost is when the coverage target is slack, and when it binds we prove that linear regret is unavoidable for every coverage-feasible algorithm. The realised stockout rate satisfies deterministically, for any noise distribution and regardless of the accuracy of the demand model, and one conformal state per segment of a fixed context partition yields the same guarantee on every segment. The analysis replaces the Lipschitz and piecewise-convexity structure of prior joint pricing and allocation arguments, which fails under atomic noise, with a coupling bound on poisson-binomial tails and a monotone-demand comparator. On synthetic benchmarks with Lipschitz and atomic noise, the fitted regret exponents match the prediction and lie far below the distribution-free baseline; when the target binds, CP-CPA attains a stockout rate of with zero variance across seeds while the plug-in quantile over-provisions at higher regret. On a demand curve calibrated from the public UCI Online Retail II log, CP-CPA is the only method that meets the target, at a profit cost of , and the per-segment variant never exceeds the target in any segment.

Primary area:
optimization
Keywords:
online learning, contextual bandits, conformal prediction, adaptive conformal inference, censored feedback, regret bounds, dynamic pricing, inventory allocation, distribution-free uncertainty quantification, safe decision making

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_rMhkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies online contextual pricing with multi-supplier inventory, price-dependent allocation and stockout constraints. The authors propose CP-CPA, which integrates residual-grid contextual pricing, profit-optimal inventory estimation, and online conformal calibration. The method aims to achieve sublinear regret while providing deterministic guarantees on the realized stockout rate.

Critical strengths
  1. The paper addresses a meaningful problem by jointly considering contextual pricing, inventory allocation, and stockout constraints.
  2. The authors claim that they realize a sublinear regret while providing deterministic guarantees on the realized stockout rate.
Critical weaknesses

The theoretical guarantees require stronger justification.

  1. The sublinear regret seems to rely on the assumption A.5. However, whether the assumption is appropriate under general atomic noise with multiple consumers still needs further discussion.
  2. The author has presented the long-term average stockout constraint. However, the linear regret lower bound in Proposition A.1 seems to depend on the safety constraints for each round.
Questions for authors
  1. Could the authors establish that Assumption A.5 holds for the realistic secnario.
  2. Could the authors clarify whether the linear regret lower bound in Proposition A.1 is true under the weaker long-term average stockout constraint?
Reviewer_o9f2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the combined setting where the decision-maker joinly determines the price and the inventory, from multiple suppliers with a known cost function. The customer valuation feedback is censored, but the purchase intention is not (so it is sort of a semi-censored feedback without the lost-sales issue). The paper proposes an algorithm that uses the first O(T2/3)O(T^{2/3}) times for uniform exploration, as done in many existing literature, to develop an estimator for the value parameter. Then it uses a grid-based pricing and inventory routine after freezing the value parameter estimation from the pilot phase. Experiments are included to help illustrate the superiority of this algorithm.

Critical strengths

The problem studied in this paper is of practical significance.

Critical weaknesses

1. Inappropriate or oversold modeling choices

  • (Hidden assumption) The most emphasized contribution of this paper is to handle the non-Lipschitz noise distribution and achieve the optimal regret. However, it makes Assumption A.5 in the appendix, which bounds how frequently the algorithm's chosen price places customers close to discontinuities in the demand survival function. Although it automatically holds under Lipschitz noise, for the non-Lipschitz case, it does basically all the weight-lifting, because the difficulty under arbitrary demand is precisely controlling regret near those discontinuities. I would say it is to some extent "hidden" because it obviously deserves to be in the main text, but it is only mentioned without explanation.
  • (Vacuous multi-suppliers) In Appendix A.1, the paper specifies that there are nn customers and mm suppliers, each with nn capacity. Then it is clear that Nt≤nN_t\le n and Kt≤nK_t\le n for every time tt. But supplier 1 has nn capacity, and the model is cheapest-supplier-first. It effectively becomes a single supplier problem, if I am not mistaken. This is also never specified in the main text and "hidden" in the appendix. The authors would need to adopt general capacities and modify the monotonicity cost for a meaningful model.

2. Proof issues

  • (Wrong Prop A.1) I am not sure Proposition A.1 is correct. Consider the following example: 2 customers; valuation is 0.30.3 or 22 with equal probability 0.50.5; 1 supplier with zero supply cost and a holding cost 0.650.65. For any inventory choice, the optimal price is clearly p=2p=2. The (reasonable) inventory choices are K=1,2K=1,2, with expected profit RK=1=0.85,RK=2=0.7R_{K=1}=0.85, R_{K=2}=0.7 and stockout probability αK=1=0.25,αK=2=0\alpha_{K=1}=0.25, \alpha_{K=2}=0. Suppose the target stockout is α=0.1\alpha=0.1. The unconstraint hindsight policy is p=2,K=1p=2,K=1, and the per-round-safe policy (as defined in Proposition A.1) is p=2,K=2p=2,K=2, which gives gap ρα=0.15\rho^{\alpha}=0.15. But on the LHS of Proposition A.1, an average-safe policy can take K=1K=1 with probability 0.40.4 and K=2K=2 otherwise, with stockout probability exactly 0.10.1 and profit 0.760.76. Extending this to TT rounds gives LHS$\le 0.09T < 0.15T=\rho^\alpha T\sim$ RHS. So somewhere the proof is wrong. Plus, I think the most natural choice is to consider the average-safe hindsight optimal in the regret benchmark, which happens to eliminate the need for such a Proposition A.1.
  • (Dropped factor) In Lemma A.11, there is an extra linear-in-TT term with sum coefficient (cm−pt)+(c_m-p_t)_+. The authors drop this term in the theoretical guarantees by saying that ptp_t is typically larger than cmc_m in the numerical experiments. And they ask readers to keep this factor in mind, even though this detail is hidden in one line of the appendix. This should never happen in any decent theory work.
  • (Invalid Lemma A.11) I don't think the first claim 0≤Ft(pt,K^t)−Ft(pt,Kt)0\le F_t(p_t,\hat K_t) - F_t(p_t,K_t) is obvious from (17), as claimed in the proof, or true at all. A nonnegative estimated marginal value does not translate to a nonnegative true marginal value. But this is minor compared to other weaknesses and seems fixable.
Questions for authors
  • Please address the weaknesses above.
  • Why don't we use the average-safety-constraint hindsight optimal policy as the regret benchmark? This is the natural choice, rather than using the unconstrained optimizer.
  • Could the authors explain more about the core challenge (after addressing the weaknesses above) faced by this paper, and the novel idea that resolves the challenge? Which part is new compared to a combination of existing techniques?
Reviewer_3sQVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies a classical online operations management problem in which a seller jointly makes pricing, inventory provisioning, and allocation decisions under uncertain demand. The objective is to maximize cumulative profit while keeping the stockout frequency below a prescribed level α\alpha. Customer valuations follow a contextual linear model with i.i.d. noise, and each customer makes a binary purchase decision. Multiple customers may arrive in each round, and stockout is measured by whether total demand exceeds available inventory, rather than by the number of unserved customers. Although multiple suppliers are considered, inventory allocation follows a relatively simple cheapest-first rule. Moreover, the seller observes individual purchase indicators and hence the total potential demand, so the model does not consider inventory-censored demand observations.

The authors propose CP-CPA, a two-stage algorithm that first performs pure exploration to estimate the valuation model and then combines optimistic pricing based on a residual-grid demand estimator, an adaptive conformal quantile for inventory provisioning, and greedy supplier allocation.

The theoretical analysis distinguishes two regimes, both using the unconstrained clairvoyant profit optimum as the regret benchmark. In the slack regime, where the profit-optimal policy already satisfies the stockout constraint, the algorithm achieves O~(T2/3)\widetilde O(T^{2/3}) regret. In the binding regime, the regret bound becomes O~(T2/3)+γmax⁡E[DT+]\widetilde O(T^{2/3})+\gamma_{\max}\mathbb E[D_T^+], where DT+D_T^+ denotes the cumulative additional inventory introduced by the conformal safety layer. This additional term is not guaranteed to be sublinear, and the paper does not establish sublinear regret against the optimal safety-constrained policy in the binding regime. Separately, the algorithm provides a deterministic finite-horizon guarantee on the realized stockout frequency.

Critical strengths

1, This paper studies a practically relevant online decision-making problem that jointly considers pricing, inventory provisioning, and stockout constraints. Its main theoretical contribution is the regret decomposition that separates the learning cost from the additional inventory cost required for safety. In particular, achieving the optimal T2/3T^{2/3} learning rate under agnostic, potentially discontinuous demand distributions is a meaningful result.

2, The deterministic stockout coverage guarantee is appealing, as it does not depend on the accuracy of the estimated demand model or the underlying noise distribution. The extension to segment-conditional coverage is also useful for applications where service levels must be maintained across different customer groups.

Critical weaknesses

I have substantial concerns about the soundness of the main theoretical results, particularly Theorem 1 and its supporting arguments.

1, The slack-regime O~(T2/3)\widetilde O(T^{2/3}) regret guarantee relies on Assumption A.6, which requires strictly positive marginal profit for all inventory units up to a demand quantile at every candidate price. However, at any sufficiently small candidate price p≤2Δp\le 2\Delta, Assumption A.3 implies that every consumer purchases, while g1(p)=p−c(1)−γ(1)<0g_1(p)=p-c(1)-\gamma(1)<0. This actually contradicts Assumption A.6, which suggests that the slack-regime guarantee could be vacuous.

2, Lemma A.11 bounds the additional inventory cost using γmax⁡+(cm−pt)+\gamma_{\max}+(c_m-p_t)_+, but the final regret bound in Theorem 1 omits the second term without establishing pt≥cmp_t\ge c_m.

3, Proposition A.1 appears incorrect because it uses a per-context safety-constrained benchmark to derive a lower bound under an average stockout constraint. In fact, policy may violate the stockout threshold on some contexts while compensating on others, potentially satisfying the average constraint with zero regret against the unconstrained oracle even when the proposed constraint gap ρα\rho_\alpha is positive.

These three points are substantive mathematical issues rather than minor gaps in presentation. They undermine the central regret guarantees and the claimed optimality interpretation.

Questions for authors

Please see what I stated above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.