acceptodds
Under review as a conference paper at ICLR 2027

Nonparametric Contextual Pricing and Inventory Learning under Censored Demand

Abstract

In online retailing, when a product sells out, a retailer often sees only the units sold, not how many customers would have bought it had inventory been available. However, the inventory level determines how much demand is revealed, and this information can influence subsequent decisions and future profits. We study an online selling problem in which, in each round, the seller observes a market context and then makes pricing and stocking decisions based on censored sales data from previous rounds. The challenge is to learn a context-dependent pricing and stocking policy without a parametric demand model or fully observed profit, since lost-sales costs depend on unobserved unmet demand. To overcome this difficulty, we propose a Surrogate-Profit UCB (SP-UCB) algorithm that turns each incomplete sales record into a reliable guide for both inventory and price decisions, using data from past rounds with similar market conditions. This design allows us to learn while serving customers, without a separate exploration phase or the need to recover all demand hidden by stockouts. We prove the minimax optimality of the proposed algorithm, with strictly faster rates when expected profit varies more smoothly with price. Comprehensive numerical experiments have been conducted to confirm the effectiveness of the proposed algorithm.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
online learning, dynamic pricing

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_teeNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies online joint pricing and inventory optimization when demand depends on contextual features and sales are censored by stockouts. The retailer observes the context, chooses a price and inventory level, and subsequently observes sales, but cannot directly observe unmet demand. Consequently, realized profit is not fully observable, making conventional reward-based online learning difficult.

The authors propose Surrogate-Profit UCB (SP-UCB), a nonparametric contextual learning algorithm that combines inventory optimization, local context pooling, and optimistic price selection. Its central idea is to construct an observable surrogate reward from sales and leftover inventory. This surrogate preserves inventory comparisons at a fixed price but has a price-dependent bias. The authors correct this bias using a low-demand quantile that can be observed by choosing inventory appropriately.

The empirical evaluation uses synthetic demand environments, varying context dimension and smoothness. The paper also reports comparisons with context-agnostic implementations of existing online pricing and inventory methods, a twice-smooth price specialization, and a smoothness-mismatch study.

Critical strengths

S1. A meaningful problem with a clear technical challenge

The paper addresses an important interaction between pricing, inventory decisions, and censored feedback. Stocking decisions affect both profitability and the information available to the learner, so the problem is more involved than standard contextual pricing with directly observed rewards.The work makes a useful distinction between learning from incomplete demand observations and simply estimating a demand curve accurately. It seeks only the information required to make good decisions.

S2. An elegant surrogate-reward construction

The two-stage calibration argument is the paper's most compelling conceptual contribution. For price comparisons, the authors use a visible low quantile to remove the price-dependent offset. Under their location-shift demand model, this correction leaves a context-dependent constant that does not affect the optimal price. This is a strong structural insight: it avoids requiring full demand reconstruction or a separate high-inventory exploration phase.

S3. Joint treatment of learning and decision-making

SP-UCB combines several elements in a way tailored to the problem:

  • Inventory profiling estimates near-optimal stocking levels.
  • Coarser inventory-search grids reuse decisions across nearby prices.
  • Kernel-weighted observations share information across nearby contexts.
  • Upper confidence bounds encourage exploration of under-sampled prices.

The use of the second-order loss from inventory estimation near the optimal stocking level is particularly useful. It allows inventory learning to be less precise than price evaluation might otherwise require, reducing the burden of joint optimization.

Critical weaknesses

W1. The demand assumptions are restrictive

The analysis assumes a location-shift structure, D=λ(x,p)+ϵ,ϵ∣x∼FxD=\lambda(x,p)+\epsilon,\qquad \epsilon\mid x\sim F_x, where the residual distribution can vary with context but does not vary with price within a context. This structure is central to the calibration argument: the low demand quantile shifts with price by the same amount as mean demand. In practical pricing applications, however, price changes can affect not only mean demand but also its variance, tail behavior, and customer composition. In those settings, the quantile correction may no longer remove the price-dependent bias.The paper acknowledges this limitation, but it remains a significant restriction on the scope of the results.

W2. Strong regularity and feasibility assumptions limit the covered settings

The model assumes smooth context dependence in both the demand location and residual distribution, bounded demand support, density bounds near relevant quantiles, and sufficient inventory capacity to keep the calibration and optimal inventory quantiles away from the action-space boundaries. These assumptions facilitate quantile estimation and inventory profiling, but they exclude potentially important cases, including irregular demand distributions, quantiles near support boundaries, and settings in which inventory capacity prevents the algorithm from exposing the calibration quantile. The analysis also assumes i.i.d. context-shock pairs and no inventory carryover.

W3. The empirical evaluation does not establish broad practical effectiveness

The experiments are synthetic, use five independent replications, and cover context dimensions of one and two over a limited set of smoothness levels and horizons.The appendix adds comparisons with two context-agnostic baselines—CWZ and a SALA-style method—on a contextual demand instance. These comparisons are informative, but they do not establish superiority over a strong contextual pricing-and-inventory method. In particular, the paper discusses Han et al. (2026a), a contextual approach using basis functions and an initial high-inventory exploration phase, but the reported comparison does not include an implementation of that method. The primary plots also show cumulative-regret scaling diagnostics rather than a direct empirical verification of minimax optimality. The authors themselves acknowledge that the fitted slopes do not establish worst-case asymptotic rates.

W4. The novelty relative to the closest prior work could be sharpened

The paper positions itself against several strands of work, particularly noncontextual censored pricing and inventory learning, contextual pricing with historical data, and the basis-function approach of Han et al. (2026a). The most convincing distinction is the combination of nonparametric contextual learning, a quantile-calibrated surrogate, and matching minimax rates without an initial phase designed to reveal demand through high inventory. However, the empirical evidence does not directly compare against the closest contextual baseline, and the paper could more explicitly separate which aspects of the contribution are fundamentally new from which extend established censored-inventory or contextual-bandit techniques.

Questions for authors

Q1. Robustness to price-dependent demand noise

The calibration identity relies on the residual distribution being independent of price within a context. Can the authors characterize a broader class of price-dependent residual distributions for which a similar observable surrogate remains valid? If not, can they explain why this restriction is essential rather than an artifact of the proposed construction?

Q2. Necessity of the quantile and feasibility assumptions

Which assumptions in Assumption 3 are required for identification of the optimal price and inventory, and which are needed only for the finite-sample regret analysis? In particular, can the method or its guarantees be extended when inventory capacity prevents the calibration quantile from being fully observable?

Q3. Comparison against contextual baselines

Why is the contextual method of Han et al. (2026a) not included in the experimental comparison? Can the authors provide a controlled comparison against this method or explain why a meaningful comparison is not possible?

Q4. Finite-horizon profiling overhead

How much of the regret is attributable to inventory profiling versus contextual price learning across the reported settings? Can the authors report profiling duration, the fraction of rounds spent profiling, and sensitivity to rare context cells?

9 comments

We thank the reviewer for the questions. Below we distinguish the assumptions' roles, establish the broader calibration condition, and provide the requested baseline and profiling evidence.

W1 and Q1. The demand restriction has established precedents, and exact calibration requires less than a location family.

Price-independent additive noise is an established specification in nonparametric pricing–inventory learning: see Petruzzi and Dada (1999), Chen, Chao, and Shi (2021, Section 2.1), and Chen, Wang, and Zhou (2024, Section 2.1). Our demand surface and context-dependent residual law are both unknown, with no finite-dimensional family imposed. Variance and shape may vary with context. Price invariance within a context restricts the class, but does not make its learning problem parametric or resolve endogenous censoring.

The broader class has an exact characterization (Appendix K.1, Eq. K.1). Write μx(p)=E[D∣x,p]\mu_x(p)=\mathbb E[D\mid x,p], let qρ,x(p)q_{\rho,x}(p) be a finite low quantile, and set O=min⁡D,yO=\min\\{D,y\\} and Y=(p+b)O−hc(y−O)Y=(p+b)O-h_c(y-O). For nonnegative, integrable demand,

=Q_x(p,y)+b\bigl[\mu_x(p)-q_{\rho,x}(p)\bigr].$$ For $b > 0$, the offset is price independent **exactly when $\mu_x(p)-q_{\rho,x}(p)=c(x)$**. Provided $y > q_{\rho,x}(p)$, this quantile is identifiable from sales. Other distributional features may depend on price. The displayed identity remains valid for general demand laws under the stated integrability conditions; arbitrary price-dependent residuals need not satisfy the price-independent-offset condition. Extending the full theorem additionally requires inventory, visibility, and estimation guarantees. **We also tested violations directly (Figure 9; complete comparison in Table 5).** Holding the nonlinear mean fixed, we used uniform residuals with radius $a_\gamma(p)=0.15[1+\gamma(2(p-0.1)/0.9-1)]$, for $\gamma\in\\{-0.75,-0.375,0,0.375,0.75\\}$. The four nonzero settings cover both dispersion directions; the extreme settings give a sevenfold standard-deviation ratio, with radii 0.0375–0.2625. At $T=16000$, with eight paired seeds and no parameter retuning, **calibration reduces true-profit regret by 6.2%–23.7% against the identical algorithm with calibration disabled, winning every paired comparison within every setting**. Thus its benefit survives these violations. The residual bias remains nonzero and empirical UCB nonvisibility is 6.5%–18.1%; this is finite-horizon evidence, not an extended regret theorem. **W2 and Q2. Identification assumptions and rate regularity must be assessed separately.** The relevant conditions serve distinct, explicit purposes: | Conditions | Role | |---|---| | Price-invariant mean-minus-quantile difference; visible calibration quantile | Identify a price-comparison signal from sales. | | Local density lower bounds; context Hölder/CDF regularity | Control quantile-estimation sensitivity and localization bias. | | Density upper bound; interior critical inventory | Make local inventory loss quadratic. | | Bounded support; capacity and quantile margins | Bound observations and ensure feasible inventories retain visibility despite estimation error. | These rate conditions have concrete precedents. [Chen, Chao, and Shi (2021, Section 2.5, Condition 4)](https://doi.org/10.1287/moor.2020.1084) require positive density near optimal quantiles, alongside strong concavity and higher-order demand smoothness. [Chen, Wang, and Zhou (2024, Assumption 1)](https://doi.org/10.1287/mnsc.2023.4859) use compact support and local density regularity. Our basic price-value class is Lipschitz, without global strong concavity. Hölder context regularity is standard in nonparametric contextual learning; see [Rigollet and Zeevi (2010, Section 2.2)](https://arxiv.org/abs/1003.1630). **We require no positive lower bound on context density.** These comparisons locate each assumption's role; they do not imply that the full model contains every earlier model. If capacity exceeds the calibration quantile but binds at the critical inventory, calibration can remain valid while our interior-inventory analysis needs modification. If capacity lies below the calibration quantile, feasible sales cannot expose that quantile, regardless of sample size. Lowering $\rho$ helps only if the new quantile is visible, regular, and satisfies the calibration relation. This is an identification constraint for the present estimator. Stationarity and no carryover define the current decision problem; temporal drift and carryover require additional analysis.

W3 and Q3. We added the contextual Han comparison and tested its dependence on the supplied representation.

Our adaptation of Han et al. (2026a, Algorithm 1) retains exploration, least-squares estimation, empirical residual estimation, and commitment. To match our problem, it removes carryover and uses feasible high inventory during exploration to reveal demand. It uses ⌈T⌉\lceil\sqrt T\rceil exploration rounds, an empirical budget choice rather than a reproduction of all theoretical constants.

In the nonlinear-interaction environment, T=16000T=16000, with five paired replications, common costs and feasible actions, and a common 801-price evaluation oracle. Han uses inventory 5 during exploration, 801 candidate prices, and features 1,p,x1−0.5,p(x1−0.5),x2−0.5,p(x2−0.5)\\{1,p,x_1-0.5,p(x_1-0.5),x_2-0.5,p(x_2-0.5)\\}.

MethodCumulative regret, mean ± SE
SP-UCB6063.19±22.066063.19\pm22.06
CWZ8642.45±170.678642.45\pm170.67
SALA8616.10±271.778616.10\pm271.77
Han-adapted, linear11653.94±451.7811653.94\pm451.78

SP-UCB reduces mean regret by 29.6% relative to the best tested comparator. CWZ and SALA pool contexts; the Han adaptation supplies the requested contextual comparator.

To separate algorithmic performance from representation choice, we also tested D=1.9+0.65cos⁡(2πx)−1.05p+ξD=1.9+0.65\cos(2\pi x)-1.05p+\xi, with independent x∼Unif[0,1]x\sim\mathrm{Unif}[0,1], ξ∼Unif[−0.15,0.15]\xi\sim\mathrm{Unif}[-0.15,0.15], T=8000T=8000, and five replications. Han uses the same estimation and commitment procedure across dictionaries, exploration inventory 2.8, and 401 candidate prices; the evaluation oracle has 801 prices.

MethodCumulative regret, mean ± SE
SP-UCB1841.28±27.021841.28\pm27.02
Han-adapted, linear3590.83±106.053590.83\pm106.05
Han-adapted, quadratic642.55±26.37642.55\pm26.37
Han-adapted, true cosine190.75±22.88190.75\pm22.88

Both experiments use p∈[0.1,1]p\in[0.1,1], b=0.4b=0.4, hc=1h_c=1, and ρ=0.2\rho=0.2. SP-UCB uses Ctune=0.01C_{\mathrm{tune}}=0.01, with cη=0.3c_\eta=0.3 and 0.250.25, respectively. Quadratic features already improve Han substantially without supplying the true cosine. A matched nonlinear mean-shape dictionary also outperforms SP-UCB in the dispersion experiment above. Performance depends materially on the supplied representation; no universal ranking follows. Our contribution is a guarantee over the stated nonparametric classes without specifying such a dictionary. Synthetic low-dimensional comparisons complement that guarantee; they do not establish broad deployment performance.

W4. Distinguishing the new contributions from established tools.

Kernel smoothing and UCB are established components. The new contribution is the quantile-calibrated price-comparison signal, its coupling with coarse inventory learning under endogenous censoring, and the matching regret bounds. The analysis jointly controls visibility, adaptive quantile estimation, localization, and profiling. It exploits locally quadratic inventory loss so that O(ε)O(\sqrt{\varepsilon}) inventory accuracy supports O(ε)O(\varepsilon) price-value accuracy. Matching upper and lower bounds establish optimal polynomial regret orders for two price-regularity classes. We will make this separation explicit in the contribution statement.

The fixed-context inventory-error appendix isolates why this matters: SP has comparison error O(e2+e′2)O(e^2+e'^2), while the analyzed CWZ Algorithm 2 correction includes a first-order term bτ(p)eb\tau(p)e. With true profit gap 0.0050.005 and errors e=−0.04e=-0.04, e′=0e'=0, the respective population gaps are 0.0062000.006200 and −0.007533-0.007533. This identifies a concrete difference in how inventory error affects price comparisons at prescribed inventories; it is not a blanket ranking of complete policies.

Q4. Profiling costs are measured explicitly; unfinished rare cells do not block learning elsewhere.

Appendix K.2 (Table 6) reports phase-regret attribution across 665 original runs, including all 150 twice-smooth runs. For each trajectory, RT=Rprof+RUCBR_T=R_{\mathrm{prof}}+R_{\mathrm{UCB}}, using expected-profit gaps on the rounds assigned to each phase. The UCB term includes residual inventory loss. Representative five-run summaries are:

SettingProfiling roundsRound share (%)Regret share (%)
Nonlinear interaction, T=16000T=160001366.80±19.831366.80\pm19.838.54±0.128.54\pm0.1225.84±0.3025.84\pm0.30
Main, d=2d=2, α=1\alpha=1, T=64000T=640003952.80±43.203952.80\pm43.206.18±0.076.18\pm0.0748.71±1.4248.71\pm1.42
Twice-smooth, d=2d=2, α=1\alpha=1, T=64000T=640003348.00±72.003348.00\pm72.005.23±0.115.23\pm0.1148.78±2.3448.78\pm2.34

Entries are means ± SE; percentage SEs are in percentage points. All cells complete profiling in these configurations. Profiling regret in the interaction experiment is 1566.42±16.081566.42\pm16.08. From T=4000T=4000 to 6400064000, the d=2d=2, α=1\alpha=1 profiling round share falls from 17.17% to 6.18% in the main experiment and from 13.09% to 5.23% in the twice-smooth experiment. Its finite-horizon regret contribution remains material. All use Ctune=0.01C_{\mathrm{tune}}=0.01; cη=0.25c_\eta=0.25 for main and 0.30.3 otherwise.

The rare-context stress test adds 45 controlled runs (Appendix K.3, Figure 10). In a fixed d=2d=2, α=1\alpha=1 interior-optimum demand environment, each coordinate independently follows w,Unif[0,0.25]+(1−w),Unif[0,1]w\\,\mathrm{Unif}[0,0.25]+(1-w)\\,\mathrm{Unif}[0,1]. We cross w∈0,0.5,0.9w\in\\{0,0.5,0.9\\} with T∈8000,16000,32000T\in\\{8000,16000,32000\\} and five paired seeds, keeping demand, costs, and algorithm parameters fixed (Ctune=0.01C_{\mathrm{tune}}=0.01, cη=0.3c_\eta=0.3).

All cells finish in all 30 runs with w=0w=0 or 0.50.5. At w=0.9w=0.9:

TTCompleted cells (%)Unfinished cells' regret share (%)Frequent-cell UCB rounds while other cells remain unfinished
800055.56±0.0055.56\pm0.002.86±0.162.86\pm0.166835.80±15.296835.80\pm15.29
1600055.56±0.0055.56\pm0.002.65±0.232.65\pm0.2313797.00±31.7713797.00\pm31.77
3200043.75±0.0043.75\pm0.004.09±0.384.09\pm0.3827223.80±35.4227223.80\pm35.42

The frequent cell (0,0)(0,0) finishes in every run. Unfinished cells account for only 0.43%–0.57% of arrivals. Local profiling allows the frequent cell to learn while others remain unfinished, whose aggregate regret share is 2.65%–4.09% here. Their per-arrival losses can still be high. The partition grows from 9 to 16 cells at the last horizon, so completion percentages concern different partitions.

These observations directly answer the concern about rare cells delaying learning. They do not establish a regret advantage over an untested synchronous policy. Under this empirical tuning, nonvisibility averages 2.31%–8.29% at w=0.9w=0.9; the experiment does not certify the theorem's visibility event. Full configuration and cell-level results are retained.

Reviewer_pR8DOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper proposes an interesting approach to perform joint pricing and inventory learning under censored demand defining theoretical regret bounds.

Critical strengths

The main strength of this paper is the definition of a surrogate profit calibrated through an observable quantile, allowing prices to be compared without reconstructing the demand tail hidden by stockouts. The mathematical proofs seem solid under the defined assumptions, and the upper and lower regret bounds are a useful contribution.

Critical weaknesses

My main reservations are related to the strength of the experimental part. The main experiments analyze the method's behavior in controlled synthetic environments, but appear to use instances whose optimal prices lie at the upper boundary of the price interval, limiting their ability to evaluate the learning of different optimal prices across contexts. The same limitation is present for the twice-smooth and smoothness-mismatch experiments.

The comparisons between online policies are limited to the appendix and use baselines that ignore context. They are also reported for a single horizon and with a small number of price points, profiler anchors, and context cells. Given the fact that the baselines ignore context, the observed advantage cannot help to distinguish the benefit derived from contextual information from that obtained through the proposed calibration and learning procedure. The authors also highlight that CWZ's performance partly reflects implementation choices, including a midpoint-inventory fallback when it reaches its search budget limit. A comparison with Han et al. (2026a), the closest contextual method, is not available.

In my opinion, the hyperparameter analysis is incomplete. The kernel used in the simulations is not explicitly specified and the tuning constants change across experiments without an analysis that separates the tuning effects from the algorithmic changes.

In retail, nonstationarity is a particularly relevant topic. The model assumes i.i.d. context–shock pairs and a time-invariant conditional demand law, so the established guarantees do not extend to drift unexplained by the observed contexts. This limitation needs to be explicitly discussed (even if it is not affecting the theoretical results).

Questions for authors
  • Could you provide the reason why you have not compared your approach with Han et al. (2026a) in a common experimental setting?

  • Could you provide more details about the kernel used in the simulations?

  • Could you define the possible direction to address the nonstationarity in future work?

7 comments

We thank the reviewer for the thorough and constructive evaluation. Your feedback helped us significantly improve our experimental comparisons and sharpen our theoretical scope. We address each of your comments and detailed questions below.

Response to Question 1: Comparison with Han et al. (2026a)

Core Design Differences and Rationale The method in Han et al. (2026a) estimates the coefficients of mean demand within a function class spanned by prescribed basis functions. In contrast, CWZ (Chen et al., 2024), SALA (Chen et al., 2021), and SP-UCB extract learning signals directly from distribution statistics. The central philosophical distinction lies in whether one must prespecify a parametric or basis function structure for mean demand. The original appendix prioritized comparisons with CWZ and SALA because they share the closest nonparametric philosophy with our work, aiming to isolate and highlight the value of the surrogate-profit construction paradigm.

We sincerely thank the reviewer for this constructive question. It made us realize that adding contextual baseline comparisons is essential for readers to clearly position our contributions. Below, we provide benchmark experiments comparing SP-UCB against an adapted version of Han et al. (2026a) under both well-specified and misspecified basis function settings.

Benchmark on the Nonlinear Interaction Demand Setting Following the setup in Appendix I.3 ("Comparisons with Neighboring Online Methods"), we evaluate the nonlinear interaction demand instance (T=16000T = 16000, 5 replications, p∈[0.1,1]p \in [0.1, 1], y∈[0,5]y \in [0, 5], b=0.4b = 0.4, hc=1h_c = 1, ρ=0.2\rho = 0.2). Each replication shares identical context and noise realizations, evaluated against the 801-point price grid oracle.

The baseline Han-adapted is implemented based on the uniform price exploration, OLS estimation, empirical residual distribution, and policy commitment framework from Section 3 (Algorithm 1) of Han et al. (2026a), using the feature dictionary 1,p,x1−0.5,p(x1−0.5),x2−0.5,p(x2−0.5)\\{1, p, x_1 - 0.5, p(x_1 - 0.5), x_2 - 0.5, p(x_2 - 0.5)\\}. To match our model setting with inventory capacity caps and zero inventory carryover:

  • The exploration inventory is set to 5 (which strictly exceeds the demand upper bound of 4).
  • Inventory carryover is removed, with optimization conducted over the feasible action space.
  • The candidate price search uses the same 801-point grid.
  • The exploration horizon is set to T0=⌈T⌉=127T_0 = \lceil \sqrt{T} \rceil = 127 as an empirical adaptation choice

SP-UCB uses cη=0.3c_\eta = 0.3 and achieves 29.6% lower mean cumulative regret than the best comparator, as summarized below.

MethodFinal Cumulative Regret (Mean ± SE)
SP-UCB6063.19±22.066063.19 \pm 22.06
CWZ8642.45±170.678642.45 \pm 170.67
SALA8616.10±271.778616.10 \pm 271.77
Han-adapted11653.94±451.7811653.94 \pm 451.78

Sensitivity to Basis Function Misspecification Assumption 1 in Han et al. (2026a) assumes that mean demand lies in the linear span of prescribed basis functions, while their Section 5.2 and Figure 3 investigate the effect of dictionary dimension. To evaluate the effect of basis function misspecification, we test a smooth nonlinear cosine demand instance:

Dt=1.9+0.65cos⁡(2πxt)−1.05pt+ξt,xt∼i.i.d.Unif[0,1],ξt∼i.i.d.Unif[−0.15,0.15].D_t = 1.9 + 0.65\cos(2\pi x_t) - 1.05 p_t + \xi_t, \quad x_t \stackrel{\text{i.i.d.}}{\sim} \text{Unif}[0, 1], \quad \xi_t \stackrel{\text{i.i.d.}}{\sim} \text{Unif}[-0.15, 0.15].

Contexts and noise shocks are mutually independent; p∈[0.1,1]p \in [0.1, 1], y∈[0,2.8]y \in [0, 2.8], with cost parameters b,hcb, h_c, and calibration level ρ\rho identical to the previous experiment. Across 5 independent runs with T=8000T = 8000, SP-UCB uses Ctune=0.01C_{\text{tune}} = 0.01 and cη=0.25c_\eta = 0.25. Han-adapted explores for ⌈T⌉\lceil \sqrt{T} \rceil rounds with inventory y=2.8y = 2.8 (fully revealing demand) and uses the same estimation and commitment steps across three feature dictionaries:

  1. Linear: 1,p,x−1/2,p(x−1/2)\\{1, p, x - 1/2, p(x - 1/2)\\}.
  2. Quadratic: adds corresponding Legendre polynomials and price interaction terms.
  3. True shape: incorporates cos⁡(2πx)\cos(2\pi x) and price interaction terms.

The candidate decision grid uses 401 points, and oracle evaluation uses 801 points.

MethodFinal Cumulative Regret (Mean ± SE)
SP-UCB1841.28±27.021841.28 \pm 27.02
Han-adapted: Linear3590.83±106.053590.83 \pm 106.05
Han-adapted: Quadratic642.55±26.37642.55 \pm 26.37
Han-adapted: True cosine190.75±22.88190.75 \pm 22.88

Under a linear dictionary, SP-UCB achieves lower regret. When provided with well-matched feature representations, the performance of Han-adapted improves significantly. This highlights that basis-function approaches depend heavily on the quality of structural priors, whereas SP-UCB learns the response surface without prespecifying parametric or basis structures.

Response to Question 2: Kernel Function Details and Robustness

We thank the reviewer for pointing this out. While the general class of admissible kernels was formally defined in the main text, the specific choice used in the simulations was inadvertently omitted from the appendix details. We are grateful for the opportunity to clarify this.

The simulation experiments implement a radial triangular kernel. Kernel Definitions and Bandwidths Let r=∥xs−x∥2/hr = \Vert{}x_s - x\Vert{}_2 / h and (a)+=max⁡a,0(a)^+ = \max\\{a, 0\\}. We evaluate three standard kernel functions:

  • Triangular: Ktriangle(r)=(1−r)+K_{\text{triangle}}(r) = (1 - r)^+
  • Epanechnikov: KE(r)=(1−r2)+K_E(r) = (1 - r^2)^+
  • Biweight: KB(r)=((1−r2)+)2K_B(r) = ((1 - r^2)^+)^2 Estimation weights are computed as ws=K(∥xs−x∥2/h)w_s = K(\Vert{}x_s - x\Vert{}_2 / h), with weighted averages and empirical quantiles normalized by Wt(xt,p)=∑swsW_t(x_t, p) = \sum_{s} w_s. The bandwidth rule is$h = \left(\frac{\log(eT)}{T}\right)^{1/(d+3\alpha)},$

and for the twice-smooth price mode, h=(log⁡(eT)/T)2/(2d+5α)h = (\log(eT)/T)^{2/(2d+5\alpha)}, where dd is the context dimension and α∈(0,1]\alpha \in (0, 1] is the input Hölder exponent.

Using cη=0.4c_\eta = 0.4 for all three kernels and holding all other parameters in the nonlinear interaction environment fixed, the empirical results across the three kernel choices are:

KernelFinal Cumulative Regret (Mean ± SE)
Triangular5413.12±14.715413.12 \pm 14.71
Epanechnikov5409.38±17.165409.38 \pm 17.16
Biweight5415.51±16.015415.51 \pm 16.01

The maximum difference across kernel specifications is approximately 0.11% of the triangular baseline, demonstrating that performance is robust to kernel choice.

Response to Weakness: Hyperparameter Tuning Analysis and Theoretical Bounds

Controlled Empirical Comparison for CtuneC_{\text{tune}} Different experimental scripts were developed at different project stages, leading to varying empirical choices for CtuneC_{\text{tune}}. We rerun the main scaling experiment across all configurations (d∈1,2d \in \\{1, 2\\}, α∈0.5,0.75,1\alpha \in \\{0.5, 0.75, 1\\}, T∈4000,…,64000T \in \\{4000, \dots, 64000\\}, 5 replications), comparing Ctune=0.05C_{\text{tune}} = 0.05 directly against Ctune=0.01C_{\text{tune}} = 0.01 under identical random seeds:

Across all six configurations and five horizons, Ctune=0.01C_{\text{tune}} = 0.01 reduces mean cumulative regret by 45.0%–70.5% relative to 0.050.05. Both choices exhibit sublinear empirical scaling over the tested horizons.

The absolute differences between fitted empirical slopes range from 0.033 to 0.115. Because CtuneC_{\text{tune}} scales the inventory profiling batch size, the interval-width stopping threshold, and the UCB exploration bonus, regret is not strictly monotonic in CtuneC_{\text{tune}}. Across tested settings, Ctune=0.01C_{\text{tune}} = 0.01 consistently achieves lower mean regret, making it a natural default.

Parameter Selection and Theoretical Certification

We thank the reviewer for prompting us to clarify the distinction between empirical tuning and theoretical certification. The revised appendix gives the guideline Ctune≤14ε,ε=hα,0<h≤1,α∈(0,1],C_{\text{tune}}\le\frac{1}{4\sqrt{\varepsilon}},\qquad\varepsilon=h^\alpha,\quad 0 < h\le1,\quad \alpha\in(0,1],which ensures Ctuneε≤ε/4C_{\text{tune}}\varepsilon\le\sqrt{\varepsilon}/4. Both 0.010.01 and 0.050.05 satisfy this bound. Full theoretical certification additionally requires the concentration and visibility conditions in Appendix C.1, with sufficiently large CtuneC_{\text{tune}} and a sufficiently small model-dependent hvish_{\text{vis}}. Certified fixed tuning preserves the regret exponent; the experiments instead use hvis=1h_{\text{vis}}=1 and assess empirical choices through expected-profit regret.

For hc>0h_c > 0, p+b>0p+b > 0 on the compact price interval P\mathcal P, and ε>0\varepsilon > 0, choosing 0<cη<min⁡p∈Pmin⁡τ(p),1−τ(p)2ε,τ(p)=hcp+b+hc,0 < c_\eta < \frac{\min_{p\in\mathcal P}\min\\{\tau(p),1-\tau(p)\\}}{2\sqrt{\varepsilon}},\qquad \tau(p)=\frac{h_c}{p+b+h_c},keeps both stockout thresholds strictly inside (0,1)(0,1); this alone does not certify quantile visibility. The static comparison uses cη=0.3c_\eta=0.3, below the resulting limit 0.34630.3463. Across five paired runs, it gives regret 6063.19±22.066063.19\pm22.06, while cη=0.4c_\eta=0.4 gives the lowest mean among the five tested values, 5413.12±14.715413.12\pm14.71. This sensitivity study quantifies the finite-horizon precision–exploration tradeoff.

Response to Question 3 & Weakness: Nonstationary Environments

Addressing nonstationarity is practically relevant in online retail (e.g., Zhao et al., 2023 for changepoint-based contextual pricing). However, integrating change detection into joint inventory and pricing under censored sales is technically demanding.

Theoretical Implications of Nonstationarity If the conditional demand law varies over time, Dt=λt(x,p)+ξt,xD_t = \lambda_t(x, p) + \xi_{t, x}, the stationarity premises in Assumptions 1–3 break. However, at any fixed period tt and context xx, as long as the residual distribution remains independent of price, has mean zero, and demand is integrable, the instantaneous calibration identity still holds:

E[Yt∣t,x,p,y]−bqρ,t(x,p)=Qt,x(p,y)−bzρ,t(x),\mathbb{E}[Y_t \mid t, x, p, y] - b q_{\rho, t}(x, p) = Q_{t, x}(p, y) - b z_{\rho, t}(x),

where Yt=(p+b)min⁡(Dt,y)−hc(y−min⁡(Dt,y))Y_t = (p+b)\min(D_t, y) - h_c(y - \min(D_t, y)), Qt,xQ_{t, x} is the realized expected profit, zρ,t(x)z_{\rho, t}(x) is the residual ρ\rho-quantile, and qρ,t=λt+zρ,tq_{\rho, t} = \lambda_t + z_{\rho, t}. For any y>qρ,ty > q_{\rho, t}, the ρ\rho-quantile of sales matches the demand quantile.

Thus, the core calibration mechanism underlying the surrogate profit does not break down due to time variation alone. However, the stationary guarantees for estimating current targets—the inventory profiling guarantee (Lemma 6), the confidence bounds (Lemmas 8–9), and the regret bounds (Theorem 1, Theorem 3, Corollary 1)—do not directly apply under time variation.

Under structured nonstationarity, theoretical guarantees can be partially restored. For example, in a piecewise-stationary setting with SS change points and known stationary segments satisfying the paper's assumptions, restarting the algorithm at each change point with segment-specific bandwidth tuning yields a dynamic regret bound of O((S+1)1−βTβ(log⁡(eT))k)O((S+1)^{1-\beta} T^\beta (\log(eT))^k) (where 0<β<1,k≥00 < \beta < 1, k \ge 0), which remains sublinear for fixed SS.

Empirical Evaluation under Nonstationarity We evaluate nonstationary drift on the nonlinear interaction environment by perturbing mean demand: λt(x,p)=λ(x,p)+at\lambda_t(x, p) = \lambda(x, p) + a_t. Defining ut=(t−1)/Tu_t = (t - 1) / T:

  • Slow periodic drift: atslow=0.1sin⁡(2πut)a_t^{\text{slow}} = 0.1 \sin(2\pi u_t) (1 full cycle over the horizon)
  • Fast periodic drift: atfast=0.1sin⁡(8πut)a_t^{\text{fast}} = 0.1 \sin(8\pi u_t) (4 full cycles over the horizon).
  • Abrupt jump: atjump=−0.2a_t^{\text{jump}} = -0.2 for ut<0.5u_t < 0.5, and +0.2+0.2 for ut≥0.5u_t \ge 0.5 (shift occurs at round 8001).
EnvironmentSP-UCBCWZSALAHan-adapted
Slow drift5705.87±25.255705.87 \pm 25.258655.89±18.428655.89 \pm 18.428810.88±290.508810.88 \pm 290.5011700.43±451.0211700.43 \pm 451.02
Fast periodic5835.04±60.975835.04 \pm 60.979269.63±26.519269.63 \pm 26.518855.93±307.018855.93 \pm 307.0111834.31±451.2011834.31 \pm 451.20
Abrupt jump8690.41±154.848690.41 \pm 154.8411659.42±319.6611659.42 \pm 319.669914.12±197.599914.12 \pm 197.599410.44±353.339410.44 \pm 353.33

We fix T=16000T = 16000 over 5 independent replications, using cη=0.4c_\eta = 0.4 for SP-UCB and otherwise identical costs, action spaces, seeds, and parameters as in the stationary setting. Algorithms (SP-UCB, CWZ, SALA, Han-adapted) are deployed without sliding windows or restarts, and evaluated against a dynamic oracle computed on the instantaneous true distribution using an 801-point grid. SP-UCB achieves the lowest mean dynamic regret across all three scenarios, reducing regret relative to the best comparator by 34.1%, 34.1%, and 7.7% under slow drift, fast periodic drift, and the abrupt jump, respectively.

Looking forward, a promising avenue is to pursue nonstationary theoretical guarantees by combining sliding windows or discounting to reduce drift bias with changepoint-triggered re-profiling to restore quantile visibility. We thank the reviewer for this constructive and thought-provoking question.

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

This paper studies online joint pricing and inventory decisions under contextual, nonparametric demand and censored sales feedback. The proposed framework consists of two parts: identification under partially unobservable profit and subsequent online learning. For identification, the authors first construct an observable surrogate that preserves expected-profit comparisons across inventory levels at a fixed price, and then use a visible low-demand quantile to remove the price-dependent bias under a context-dependent location model with price-invariant residual noise. For online learning, SP-UCB combines inventory profiling, low-quantile estimation, and kernel-based contextual UCB, with matching minimax regret bounds established for Lipschitz and twice-smooth price-value classes.

Critical strengths
  1. An elegant identification mechanism under censored demand.

Under censored demand, the unobserved lost-sales penalty makes realized profit partially unobservable, complicating comparisons across inventory levels and prices. The authors first construct an observable surrogate that preserves expected-profit comparisons across inventory levels at a fixed price, and then use a low-demand quantile made visible through inventory profiling to remove the price-dependent bias. Their combination provides an elegant and novel way to address the partially unobservable profit.

  1. Beyond Standard Kernel-UCB: Additional Technical Analysis with Matching Minimax Guarantees.

The authors develop an end-to-end learning algorithm and establish matching regret upper and lower bounds for two price-regularity classes. Beyond the standard kernel-UCB components, the analysis addresses additional technical challenges specific to censored feedback, including controlling inventory-profiling costs, establishing concentration guarantees for adaptive low-quantile estimation, and accounting for errors arising from approximately optimal inventories.

Critical weaknesses
  1. The central identification mechanism and subsequent analysis rely on a restrictive priceinvariant residual distribution assumption.

The low-quantile correction, which is central to the paper’s identification of optimal prices from censored sales, critically relies on the assumption that, conditional on context, changing the price shifts the demand distribution without altering its residual distribution. This assumption is considerably more restrictive than the general nonparametric-demand formulation may initially suggest. In many natural demand models, prices affect not only expected demand but also its dispersion and potentially its distributional shape. For example, binomial and Poisson demand models naturally exhibit price-dependent variance as purchase probabilities or arrival rates change with price.

This limitation is consequential because it directly affects identification rather than merely simplifying the analysis. When the residual low quantile varies with price, the corrected surrogate retains a price-dependent bias that can change the ordering of prices, even with arbitrarily accurate estimation. Consequently, the subsequent UCB analysis need not provide regret guarantees relative to the true profit objective.

This structural dependence substantially limits the applicability and robustness of the proposed learning framework. Moreover, the experiments do not directly examine its sensitivity to price-dependent residual distributions.

  1. Limited methodological novelty beyond the identification mechanism.

The proposed method largely combines a surrogate-based identification mechanism with relatively standard contextual kernel-UCB learning. Beyond this mechanism, the algorithmic design and regret analysis rely substantially on established techniques, leaving the incremental methodological contribution somewhat limited.

Questions for authors

Q1. The role of the unmet-demand penalty bb.

The profit function includes a per-unit unmet-demand penalty b(D−y)+b(D-y)^+, which makes realized profit partially unobservable under demand censoring. When b=0b=0, realized profit becomes directly observable despite demand censoring.

(a) Economic motivation.

I find this additional penalty somewhat confusing, since unfulfilled sales already reduce realized revenue. What is the economic interpretation of this additional penalty? How is this parameter determined in practice, and what motivates treating it as known?

(b) Implications for identification and learning.

Could the authors clarify how the main learning difficulty and methodological contribution change when b=0b=0? Conversely, a positive bb increases the critical-fractile level of the optimal inventory, which can facilitate low-quantile visibility. How important is this dual role of bb to the proposed identification and learning framework?

Q2. Dependence on price-invariant residual distributions.

As discussed in Weakness 1, the low-quantile calibration relies on the residual demand distribution being independent of price conditional on context. Violating this assumption may introduce a price-dependent bias that distorts the true profit ranking.

(a) Algorithmic and theoretical challenges.

What additional identification and learning challenges would arise if the residual distribution were allowed to depend on price? What modifications to the proposed methodology and theoretical analysis would be needed, and how substantial would these challenges be?

(b) Practical robustness.

Could the authors provide numerical experiments with price-dependent residual dispersion to assess the robustness of the proposed method under such model misspecification?

Q3. The role of censoring in minimax learning difficulty.

The lower-bound construction appears to reflect the intrinsic difficulty of nonparametric contextual price learning, even under full-demand feedback. Does this suggest that censoring introduces no additional polynomial-order minimax regret penalty under the stated assumptions? If so, could the authors clarify why this is possible and how censoring affects the fundamental learning difficulty of the problem?

3 comments

We thank the reviewer for the thoughtful questions. We clarify the assumptions and contribution below, supported by a broader calibration identity and new controlled experiments.

Q1(a). The shortage penalty represents an additional economic cost.

Foregone revenue is already captured by pmin⁡D,yp\min\\{D,y\\}. The term b(D−y)+b(D-y)^+ represents an additional shortage cost, such as service compensation or damaged customer goodwill. This distinction follows the classical formulation of Petruzzi and Dada (1999, Section 1); Chen, Wang, and Zhou (2024, Eq. (1)) use the same objective and goodwill interpretation. Known bb is a supplied business cost defining the objective, while demand remains unknown. Estimating goodwill costs jointly with demand is a different problem. We will clarify this interpretation.

Q1(b). With b=0b=0, profit is observable; positive bb can help quantile visibility.

When b=0b=0, realized profit is pO−hc(y−O)pO-h_c(y-O), so our correction for hidden shortage costs becomes unnecessary. Price and inventory learning remain necessary; our formal theorems concern fixed b,hc>0b,h_c > 0.

The reviewer is correct that ϕ(p)=(p+b)/(p+b+hc)\phi(p)=(p+b)/(p+b+h_c) increases with bb. Holding demand fixed, a higher penalty raises the optimal inventory quantile and can increase its separation from the calibration quantile. This helps visibility when capacity and quantile margins permit it; positive bb alone does not guarantee visibility.

Q3. Yes: censoring causes no additional polynomial minimax penalty within our stated classes.

The lower bound already holds with full-demand feedback: its information bound is established before applying the censoring map. Conversely, a full-demand learner can emulate a censored policy. Combined with our upper bounds, this gives the same polynomial minimax orders under both feedback models:

\qquad \widetilde\Theta\\!\left(T^{\frac{2d+3\alpha}{2d+5\alpha}}\right).$$ A visible low quantile supplies the required price comparison without reconstructing full demand. Locally quadratic inventory loss means $O(\sqrt{\varepsilon})$ inventory accuracy suffices for $O(\varepsilon)$ price-value accuracy, making profiling lower order at the selected bandwidths. Censoring still incurs finite-horizon profiling costs. Its structure-dependent cost is consistent with [Besbes and Muharremoglu (2013)](https://doi.org/10.1287/mnsc.1120.1654); our matching bounds establish the conclusion for contextual joint pricing and inventory learning. **W2. The contribution is constructing and analyzing the censored-feedback signal required by UCB.** Standard contextual UCB does not itself establish quantile visibility, remove hidden price-dependent costs, or determine the required inventory precision. Our analysis couples local profiling, adaptive quantile estimation, and coarse inventory/fine price learning, with matching lower bounds for two price-regularity classes. The inventory-error appendix isolates the distinction: the SP population comparison error is $O(e^2+e'^2)$, whereas the analyzed CWZ Algorithm 2 correction contains an additional first-order term $b\tau(p)e$. This explains the different inventory-precision requirements at the specified inventories; it is not a claim that the full CWZ policy must fail.

W1 and Q2(a). Price-independent noise is an established modeling choice; the calibration identity applies more broadly.

Additive uncertainty is classical in Petruzzi and Dada (1999), and price-independent additive noise is used in Chen, Chao, and Shi (2021, Section 2.1) and Chen, Wang, and Zhou (2024, Section 2.1). Our unknown mean surface and residual distribution require no finite-dimensional specification; residual variance and shape may vary with context. The restriction concerns price dependence at a fixed context. We will make this scope explicit.

More generally, let μx(p)=E[D∣x,p]\mu_x(p)=\mathbb E[D\mid x,p], let qρ,x(p)q_{\rho,x}(p) be a finite low quantile, and set Y=(p+b)O−hc(y−O)Y=(p+b)O-h_c(y-O). For nonnegative, integrable demand,

=Q_x(p,y)+b\bigl[\mu_x(p)-q_{\rho,x}(p)\bigr].$$ For $b > 0$, the offset is price independent **if and only if $\mu_x(p)-q_{\rho,x}(p)=c(x)$**. Thus location shifts are sufficient; invariance of the entire residual distribution is unnecessary for this identity. When $y > q_{\rho,x}(p)$, the required quantile is identifiable from sales. Arbitrary price-dependent residuals can violate this condition and distort price rankings. An extension must identify the remaining correction and control its estimation, inventory error, and visibility. The broader identity alone does not establish an extended regret theorem. The experiment below directly tests whether calibration remains useful when the condition fails. **Q2(b). Calibration retains its benefit under the tested price-dependent dispersion.** Appendix K.1 (Figure 9 and Table 5) reports the new controlled experiment. We hold the nonlinear contextual mean fixed and replace the residual by $a_\gamma(p)U$, where $U\sim\mathrm{Unif}[-1,1]$ and $$a_\gamma(p)=0.15\left[1+\gamma\left(2\frac{p-0.1}{0.9}-1\right)\right].$$ The four nonzero settings below violate price invariance in both directions; zero is the control. At $|\gamma|=0.75$, standard deviation varies sevenfold across prices, with absolute radii 0.0375–0.2625. Each setting uses $T=16000$ and eight paired seeds, with fixed costs, actions, and algorithm parameters and no retuning. Learning uses censored sales; regret uses true expected profit, including shortage costs. The ablation changes only the calibration switch. | $\gamma$ | SP-UCB regret | Without calibration | Reduction | |---:|---:|---:|---:| | $-0.75$ | $5888.0\pm24.6$ | $7716.0\pm31.5$ | 23.7% | | $-0.375$ | $5920.6\pm30.7$ | $7485.3\pm31.4$ | 20.9% | | $0$ | $6034.2\pm29.4$ | $7258.7\pm26.5$ | 16.9% | | $0.375$ | $7402.0\pm16.2$ | $8237.6\pm26.2$ | 10.1% | | $0.75$ | $7567.1\pm17.3$ | $8068.9\pm22.3$ | 6.2% | Entries are means ± SE. **Calibration improves regret in all eight paired seeds in every setting**, including all four misspecified settings. Its benefit narrows as dispersion increases with price, quantifying the sensitivity. The mechanism is explicit: calibration replaces the mean-demand offset $b\lambda(x,p)$ with $b(1-2\rho)a_\gamma(p)$. With $b=0.4$, $\rho=0.2$, and optimal inventories, a separate population diagnostic on a $41\times41$ context grid gives average decision loss below $2.42\times10^{-4}$ for the calibrated score, versus 0.0232–0.0324 without calibration. These results establish a finite-horizon calibration benefit under the tested violations. Residual bias remains nonzero, and empirical UCB nonvisibility ranges from 6.5% to 18.1% (10.9% in the control); they do not extend the theorem. Table 5 retains all five tested methods.

I thank the authors for their detailed response. It clarifies the modeling assumptions and provides additional evidence on robustness to price-dependent noise. I will keep my score unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.