Nonparametric Contextual Pricing and Inventory Learning under Censored Demand
Zean Han, Jing Liang, Ruihan Lin, Zezhen Ding, Jiheng Zhang
cs.LG
Quick verdict
声明了基于 MCK-UCB 的、具有匹配边界的删失销售（censored sales）上下文最小极小极大最优定价-库存学习；证明详尽，但实验仅限于合成数据。
Summary
本文研究了一位在线卖家，其在每一轮中观察到一个市场情境（market context），并且必须同时选择价格和库存水平，同时仅从因缺货（stockouts）而被删失（censored）的销售记录中学习：每当需求超过库存时，真实需求和利润中的缺货损失部分将保持隐藏状态。所提出的 MCK-UCB 算法建立在一个可观察的销售代理 Y = (p+b)O − hc(y−O)（等于真实利润加上需求项 bD）以及分位数校准（quantile calibration）的基础上，该分位数校准消除了价格相关的偏差，从而保留了上下文内的价格排序。库存通过在粗略的上下文单元和粗略的价格锚点上根据缺货指标进行二分（bisection）来配置，而价格则在更细的决策网格上学习，使用的是汇集了邻近上下文的核估计。作者证明了匹配的最小极大遗憾率（minimax regret rates）——对于 Lipschitz 价格值，率为 Õ(T^((d+2α)/(d+3α)))；对于二阶平滑价格值，率为 Õ(T^((2d+3α)/(2d+5α)))——并提供了基于打包（packing-based）的下界。合成实验（视界高达 100,000；d ∈ {1,2}；α ∈ {0.5,0.75,1}；五次重复）显示遗憾率接近预测指数，呈亚线性增长，此外还进行了平滑度失配诊断和二维比较（T=8000，八次重复），其中 MCK-UCB 的平均遗憾率 2570.1 优于非上下文基线 CWZ (4957.7) 和 SALA (5482.1)。主要局限性：单周期易腐烂库存模型、跨价格共享噪声限制、已知的平滑度 α，以及没有真实数据或已发布代码的纯合成环境。
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An observable sales proxy Y=(p+b)O−hc(y−O) satisfies the sample-path identity Y = r + bD, so censored sales alone can steer inventory decisions without observing demand or realized profit (Proposition 3).
- Subtracting an estimable visible-quantile term b·q̂ρ(x,p) from the proxy yields a calibrated value whose within-context price maximizers coincide with those of the true optimized profit Gx(p), because the residual offset b·zρ(x) is common to all prices (Lemma 1).
- MCK-UCB achieves minimax regret Θ-tilde(T^((d+2α)/(d+3α))) under Lipschitz price values and Θ-tilde(T^((2d+3α)/(2d+5α))) under twice-smooth price values, with matching lower bounds within the stated shared-noise class (Theorems 1-3, Corollary 2).
- Two-resolution asynchronous inventory profiling costs Õ(h^(−d/2)·ε^(−3/2)) rounds, stays lower-order at the optimal bandwidths, and makes the calibration quantile visible with O(ε) value loss (Lemma 2, §4).
- In the two-dimensional synthetic comparison (T=8000, eight repetitions, common random numbers), MCK-UCB attains mean cumulative regret 2570.1 versus 4957.7 (CWZ) and 5482.1 (SALA).
Strengths
Clean identification device for censored feedback
Proposition 3 proves the sample-path identity Y(p,y,O) = r((p,y),D) + bD for the observable proxy Y=(p+b)O−hc(y−O), so the proxy preserves inventory maximizers exactly; Lemma 1 then shows calibration by a visible ρ-quantile leaves a within-context common shift, so price rankings match the true optimized profit Gx(p) without ever recovering censored demand.
Complete, matching minimax theory
Theorem 1/Corollary 2 give Θ-tilde(T^((d+2α)/(d+3α))) Lipschitz and Θ-tilde(T^((2d+3α)/(2d+5α))) twice-smooth rates, with a full proof chain: supermartingale concentration handling data-dependent kernel weights (Lemma 3, §F), deterministic counting (Lemma 5), and a packing-based direct-sum lower bound (Definition 1, Lemmas 6-8, §I) where the hard subfamily even reveals Fx≡F0 to the learner.
Sharp positioning against prior work
Unusually disciplined experimental protocol
Honest scope statements
Weaknesses
Degenerate noise family in all experiments
Baselines are noncontextual by design
Ambiguous exponent reporting in Fig. 1
Non-adaptive tuning requirements
Operational limits of the inventory model
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper studies an online seller who, each round, observes a market context xt ∈ [0,1]^d and jointly chooses a price pt and inventory yt, learning only from censored sales Ot = min{Dt, yt} with Dt = λ(xt, pt) + ϵt. Profit itself is unobservable in stockout rounds because the shortage term b(D−y)+ depends on hidden demand. Three claims are made: (1) an observable learning signal exists under censoring; (2) the MCK-UCB algorithm learns jointly without a separate exploration phase; (3) minimax regret Θ-tilde(T^((d+2α)/(d+3α))) for Lipschitz price values and Θ-tilde(T^((2d+3α)/(2d+5α))) for twice-smooth price values, with matching lower bounds (Theorem 1, §4). Scope is honestly delimited: Assumption 3 restricts residual laws to be shared across prices within a context (Remark 1), inventory is single-period perishable with no carryover (§2, §7), and bandwidths require known T, d, α (§B.1).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is censored demand in pricing and inventory problems?
When a product sells out, the seller observes only the number of units sold, not the customers who would have bought at higher stock. True demand and the lost-sales cost are then hidden exactly when demand is strongest. This paper shows how to learn both price and stocking decisions online from such truncated sales records, using an observable proxy equal to profit plus a demand term, plus a quantile calibration.
What regret guarantees does the MCK-UCB algorithm provide?
The paper proves matching (minimax-optimal) regret rates up to logarithmic factors: Õ(T^((d+2α)/(d+3α))) when the optimized profit is Lipschitz in price, improving to Õ(T^((2d+3α)/(2d+5α))) when it is twice differentiable with bounded second derivative, where d is the context dimension and α the Hölder smoothness across contexts.
Is code or data available for this paper?
The provided text contains no code or data release statement. It does give unusually complete reproduction material: full pseudocode (Algorithms 1-2), all environment parameters (e.g., b=0.4, hc=1, ρ=0.2, Ctune=0.1), explicit synthetic demand surfaces, and deterministic seed formulas in Appendix C. All experiments are synthetic; no real retail data is used.
How does this differ from earlier censored pricing-inventory papers like Chen et al. 2021/2024?
Chen et al. (2021, 2024) handle joint pricing and inventory with censored demand but no market context — one unchanging demand environment. This paper adds an observed context (season, product, customer segment) under a nonparametric demand model, so the algorithm must pool information only across nearby contexts, and it proves the first matching minimax rates for that online contextual censored-sales setting.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- PDF extraction flattens mathematical notation: sub/superscripts render inline (e.g., 'Gx(p)' for G_x(p), 'eΘ' for Θ-tilde) and in-text equation references appear as 'Section' placeholders; this affects formatting only — the mathematical content is fully readable.
- Figures are available only via captions and numbers quoted in the text; the plotted curves themselves could not be inspected.
- The text runs complete from abstract through the full lower-bound proof (Appendix I); no end-of-document truncation was observed.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.