Doubly Robust Distributionally Robust Offline Contextual Pricing

1. Introduction

• Logged observational OPE/L can be more practical and safer than real-time exploration when many price options are possible.Personalized pricing accounts for consumer heterogeneity through individual contextual information.Source: §1 · p. 1
• The framework uses a scaled kernel for continuous pricing and handles discontinuous revenue through s-smoothness of the demand-noise CDF for s=1,2.Pricing revenue is discontinuous because purchases follow a threshold rule.Source: §1 · p. 2
• The localized DR evaluation estimator achieves semiparametric efficiency under DRO while fitting nuisances only three times.It is designed to remain robust to environment shifts and estimation errors.Source: §1 · p. 2
• The proposed OPL method achieves a finite-sample regret bound of ~O_p(T^-s/(2s+1)) for s=1,2, improving on the O_p(T^-1/4) rate in prior DRO and non-robust continuous-treatment approaches.The bound holds for policy classes with finite pseudo-dimension, even when nuisance functions are estimated nonparametrically at slow rates.Source: §1 · p. 2

1.1. Related Work

• Continuous-treatment DRO OPL work has used kernel-smoothed IPW, whose performance depends on propensity-density convergence and can deteriorate under nonparametric estimation.The paper contrasts this sensitivity with its doubly robust approach, which achieves efficiency under milder product-rate conditions and a sharper regret bound [Leung et al. (2025)].Source: §1 · pp. 1–2, §1.1 · p. 2

2.1. Offline Contextual Pricing

• Table 1 compares OPL regret upper bounds by sample size T, treating policy-class complexity measures as constant factors.For the paper and Ai et al. (2026), s=1,2 denotes outcome-function smoothness; Colangelo & Lee (2026) and Leung et al. (2025) require three-times differentiability of conditional outcome densities.Source: Table 1 · p. 3
• Offline data contain contexts, prices selected by a logging policy, and binary purchase outcomes that determine observed revenue.The buyer's valuation is modeled as f(X)+Z, and the purchase decision is threshold-based.Source: §2.1 · pp. 2–3
• Overlap requires the logging policy to assign every feasible price a nonzero probability at each context.This supports evaluation of new pricing strategies using the training data.Source: §2.1 · p. 3
• The assumptions bound customer acceptance and rejection probabilities away from zero and one, and require the noise CDF to have bounded derivatives of order s.The paper discusses s=1 and the stronger s=2 smoothness case.Source: §2.1 · p. 3

2.2. Distributionally Robust Performance

• A policy maps contexts to feasible prices, and its distributionally robust regret is the difference between the optimal robust value and its own robust value.The analysis assumes the maximum is attainable and that robust value is positive, which requires the perturbation budget not to be overly large.Source: §2.2 · pp. 3–4
• The distributionally robust framework evaluates a policy against distributions in an uncertainty set centered at the nominal training distribution, with radius δ.The uncertainty set is defined using KL divergence.Source: §2.2 · p. 3
• The robust policy value admits a dual characterization with a unique maximizing scalar α, reducing the uncertainty-set optimization to a scalar dual problem.The dual representation uses W(π, α), an expectation involving exponentiated negative revenue scaled by α.Source: §2.2 · pp. 3–4
• Kernel smoothing replaces exact treatment matching for continuous prices, while doubly robust methods address the unknown logging policy and the lack of semiparametric efficiency of IPW variants.The policy value is defined through expected revenue under the nominal distribution.Source: §2.2 · pp. 3–4

3.1. Localized Doubly Robust Estimation

• The evaluation target is characterized through the dual parameter α and the moments W0 and W1, with robust value obtained from their dual-form relationship.The target moments are defined as expectations involving exponentiated negative revenue.Source: §3.1 · p. 4
• Kernel smoothing makes continuous policy evaluation feasible, and the estimator uses a doubly robust moment equation because counterfactual outcomes are unobserved.The moment equation uses observed context, price, and purchase variables with nuisance functions.Source: §3.1 · p. 4, §3 · p. 4
• Localization fits outcome nuisances near a preliminary estimate of the optimal dual parameter instead of across the full α range, requiring only three machine-learning models.This addresses the continuum of regressions created by the coupling between outcome regressions and α.Source: §3.1 · pp. 4–5
• The procedure uses cross-fitting to estimate the propensity and outcome nuisances, then solves an empirical moment equation for the dual parameter and returns the robust value estimate.The evaluation algorithm is tailored to continuous actions using kernel smoothing.Source: §3.1 · p. 5

3.2. Asymptotic Analysis

• The estimator achieves semiparametric efficiency under DRO, with a product-rate condition on nuisance estimation that is weaker than the 1/2 threshold for root-T consistency in discrete-action settings.Its covariance depends on conditional variance rather than the second moment used by self-normalized IPW, even when the logging policy is known.Source: §3.2 · p. 5
• Asymptotic normality requires undersmoothing so the kernel-approximation bias vanishes, enabling valid frequentist confidence intervals without explicit bias correction.The undersmoothing choice yields a regret rate slightly slower than ~O_p(T^-s/(2s+1)).Source: §3.2 · pp. 5–6

4.1. Learning Algorithm

• Local weighting estimates the outcome-regression continuum across α values, avoiding a separate regression-model fit for every dual parameter.The learning procedure uses forest-induced weights to estimate m0(x, p; α).Source: §4.1 · p. 6
• Alternating optimization updates the dual parameter α and policy π because their joint optimization is generally non-convex and computationally intractable.The procedure is computationally efficient but guarantees convergence only to a local maximum.Source: §4.1 · p. 6, §4.2 · p. 7

4.2. Statistical Performance Guarantee

• The guarantee requires uniform L∞ convergence of nuisance estimators across the continuous treatment manifold.The paper identifies this as a theoretical cost of ensuring robustness across that manifold.Source: §4.2 · p. 7
• With high probability, learned-policy distributionally robust regret is bounded by ˜O_p(T^-s/(2s+1)) at bandwidth h=Θ(T^-1/(2s+1)); for s=2, this sharpens the O_p(T^-1/4) continuous-action IPW bound.The rate requires γπ + γm ≥ s/(2s+1) in Assumption 4.2; the comparator uses h=Θ(T^-1/8).Source: §4.2 · p. 7

5.1. Simulation Experiment

• CDR2O2PL-CP achieves the highest robust values and slightly tighter confidence intervals than DRO-IPW at large sample sizes in simulation.The non-robust DR baseline has substantially wider intervals, while both robust methods produce more stable estimates.Source: §5.1 · p. 8
• For OPE, LDR2O2PE-CP has faster MSE convergence and outperforms DRO-IPW at sufficiently large sample sizes, but may be less stable below T=1000.Figure 1 compares MSE across training sizes, bandwidth choices, and test environments.Source: Figure 1 · p. 8, §5.1 · p. 8

5.2. Empirical Experiment - The Expedia Study

• The learned Expedia linear policy outperforms both the non-robust linear policy and the DRO-IPW policy.The study trains on hotel listings rated one to three stars and tests on four- and five-star listings.Source: §5.2 · p. 9

Conclusion

• Localized evaluation fits only three regressions, remains computationally tractable like standard non-robust DR, and achieves semiparametric efficiency.The framework is designed to protect against shifts across environments.Source: §6 · p. 9
• The learning method achieves a finite-sample regret bound of Õ_p(T^-s/(2s+1)) for s = 1, 2, improving on prior DRO rates and aligning with non-robust approaches.It uses data-driven weights across a continuum of regressions and leverages demand-noise smoothness to handle discontinuous purchase outcomes.Source: §6 · p. 9
• Comprehensive simulations show robustness to distributional shifts.The paper uses KL divergence and identifies Wasserstein distance as a direction for future work.Source: §6 · p. 9
• Figure 2 compares CDR2O2PL-CP with non-robust DR and DRO-IPW across bandwidth choices and test environments, plotting robust policy value against training sample size.The caption specifies δ ∈ {0.1, 0.2, 0.3}, δ_train = 0.2, and 2500 test samples.Source: Figure 2 · p. 9

Impact Statement

• The framework is intended to enhance the reliability of data-driven decision-making under environmental shifts.The statement names resource allocation and economic policy analysis as example domains.Source: Impact Statement · p. 10
• The work primarily aims to contribute to methodological progress in machine learning and causal inference.Source: Impact Statement · p. 10
• For socially sensitive uses such as dynamic pricing, practitioners are encouraged to audit for discriminatory outcomes or unfair treatment of subpopulations.Source: Impact Statement · p. 10

A. Proof of Results in Section 2

• The proof invokes strong duality by Theorem 1 of Hu & Hong (2013).Source: A. Proof of Results in Section 2 · p. 14
• Cauchy–Schwarz gives concavity of φ(π, α), and the non-constant reward condition makes it strictly concave for α > 0.The proof notes equality in Cauchy–Schwarz only when reward is almost surely constant.Source: A. Proof of Results in Section 2 · p. 14
• When Rδ(π) > 0, the optimal dual parameter satisfies α*(π) ∈ (0, ᾱ].The proof uses the bounds on φ(π, α) and its limit as α approaches zero to rule out α*(π) = 0.Source: A. Proof of Results in Section 2 · p. 14
• The condition δ < −log ρ is sufficient for Rδ(π) > 0, as shown by evaluating the derivative limit as α approaches zero.Source: A. Proof of Results in Section 2 · p. 14

B. Proof of Results in Section 3

• Assumption B.2 supplies Jacobian and covariance stability conditions, while its localized differentiability condition supports efficiency using a preliminary dual-parameter estimate.The proof describes B.2 as adapted to continuous pricing.Source: B. Proof of Results in Section 3 · p. 15
• As h → 0, the moment condition becomes first-order insensitive to nuisance-estimation errors; s = 1 under Lipschitz continuity and s = 2 under a second-order bounded derivative.Source: B. Proof of Results in Section 3 · p. 15
• The proof organizes the asymptotic analysis around localization and nuisance-estimation error, Jacobian stability, and an empirical-process linear representation.Source: B. Proof of Results in Section 3 · p. 15
• The estimator has an asymptotically linear representation, with its limiting distribution obtained by the central limit theorem.Source: B. Proof of Results in Section 3 · p. 21
• The efficiency of W0* and W1* propagates to the parameter vector θ* through Hadamard differentiability, yielding the semiparametric efficiency lower bound for its components.Source: B. Proof of Results in Section 3 · p. 21

Proof of Theorem 4.4

• Theorem 4.4’s proof combines Lemmas C.2, C.4, and C.5 to establish its result for s=1,2 with probability at least 1 − 7ϵ.The supplied passage does not include the bound’s explicit expression.Source: C. Proof of Results in Section 4 · p. 22

Lemma C.2

• Lemma C.2 bounds the difference between the robust risk and an infeasible estimator at order h^s for s=1,2 under Assumptions 2.1, 2.3, 2.4, and 3.2.For the first-order bounded-derivative case, the proof gives an O(h) bound.Source: C. Proof of Results in Section 4 · pp. 23, 25

Lemma C.4

• Lemma C.4 provides a high-probability bound under Assumptions 2.2 and 3.2, with probability at least 1 − 3ϵ.The supplied passage states the probability and assumptions but omits the bound’s expression.Source: C. Proof of Results in Section 4 · p. 27

Lemma C.5

• Lemma C.5 gives a high-probability bound under Assumptions 3.2, 2.2, and 4.2 for s=1,2, with probability at least 1 − 4ϵ.The stated result also requires a bandwidth condition, but its threshold is missing from the supplied passage.Source: C. Proof of Results in Section 4 · p. 31

• The proof completes the stated result for smoothness orders s = 1, 2.Source: C. Proof of Results in Section 4 · p. 37
• The complexity proof uses pseudo-dimension-based covering bounds in place of VC-dimension or exponential-covering-number assumptions used in prior arguments.The passage notes this differs from the final inequality in the proof of Lemma A.5 in Kitagawa & Tetenov (2018).Source: C. Proof of Results in Section 4 · pp. 37–38
• Under Assumption 3.2, the estimated normalization term is at least ρ/2 with probability at least 1 − ϵ.Source: C. Proof of Results in Section 4 · p. 39

D. Proof of Auxiliary Results

• Covering and packing numbers are defined through smallest covers and largest separated subsets, with empirical and uniform versions specified for function classes.The appendix relates packing and covering numbers and defines empirical distances used in these quantities.Source: D. Proof of Auxiliary Results · p. 40
• Pseudo-dimension is linked to VC dimension for thresholded function classes, with covering-number bounds for classes of bounded real-valued functions.The appendix also gives packing-number results for finite-pseudo-dimension classes.Source: D. Proof of Auxiliary Results · p. 40
• Pseudo-dimension is unchanged by nonzero affine transformations, grows at most proportionally under Lipschitz composition, and at most doubles under multiplication by a fixed function.The proof uses shattering and covering-number arguments to establish these stability properties.Source: D. Proof of Auxiliary Results · p. 40