acceptodds
Under review as a conference paper at ICLR 2027

Robust Pricing Policy Improvement with Invalid Instrumental Variables

Abstract

We study offline policy improvement over a continuous price space when historical prices are endogenous and instrumental variables (IVs) are potentially invalid. Standard IV-based methods rely on instrument exogeneity. When this assumption is violated, the demand function cannot be identified, potentially leading to suboptimal pricing policies. To address this challenge, we relax the exact exogeneity assumption by constructing an ambiguity set of demand functions whose deviations from instrument exogeneity are bounded by a radius. We propose a robust policy that maximizes the worst-case expected revenue improvement over a baseline policy across this set. To overcome the computational challenge of the resulting max-min optimization, we derive a closed-form solution for the inner minimization problem. Furthermore, we develop a safe policy improvement procedure that maximizes the allowable violation radius of the ambiguity set subject to a prespecified worst-case revenue improvement. Theoretically, we establish convergence rates for the max-min policy regret and show that the proposed safe policy achieves the target revenue improvement up to a vanishing error. Simulation studies and synthetic dataset analysis in the context of airline-ticket pricing demonstrate the superior performance of our approach under instrument invalidity compared with existing methods.

Primary area:
learning theory
Keywords:
Invalid Instrumental Variables, Robust Policy Improvement, Personalized Pricing

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_fu5DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies offline pricing policy optimization under endogenous prices and potentially invalid instrumental variables. The authors propose a set of plausible demand functions and optimize worst-case revenue improvement relative to a baseline. The authors also introduce a safe policy improvement procedure and establish theoretical regret guarantees. The approach is evaluated through synthetic simulations.

Critical strengths

I find the problem to be interesting and relevant for many e-commerce applications. Relaxing the standard IV exogeneity assumption is well-motivated, the proposed optimization formulation is intuitive. I also appreciate having the formal regret bounds.

Critical weaknesses

My main concern about the paper is that all evaluations seem to rely on synthetic data only. The main simulations use relatively simple demand functions and low-dimensional feature spaces. While these experiments demonstrate the method's behavior in controlled settings, they provide very limited evidence of how it would perform in a more realistic real-world setting. Experiments with real data, more complex demand functions, and weaker instruments would substantially strengthen the paper. Also it would probably help to consider other pricing baselines that are used by the practitioners.

It also seems to me that the choice of \rho is central to the method - and it is still unclear to me how to select this parameter in practice when the degree of instrument invalidity is unknown. In the main simulations, \rho is chosen using the true violation magnitude, which if I understood this correctly would generally be unavailable in practice. A more systematic robustness analysis for this parameter and its misspecification could be helpful.

Questions for authors

(see all my questions above)

8 comments

We appreciate the reviewer's careful reading and constructive suggestions. Below, we address each of your comments point by point.

Weakness 1: My main concern about the paper is that all evaluations seem to rely on synthetic data only. The main simulations use relatively simple demand functions and low-dimensional feature spaces. While these experiments demonstrate the method's behavior in controlled settings, they provide very limited evidence of how it would perform in a more realistic real-world setting. Experiments with real data, more complex demand functions, and weaker instruments would substantially strengthen the paper. Also it would probably help to consider other pricing baselines that are used by the practitioners.

Response to Weakness 1:

Thank you for this insightful comment.

We added two complementary experiments: supplementary simulations with 20 covariates, nonlinear heterogeneous demand, endogenous prices, and misspecified working demand models; and an application using real breakfast cereal sales data. The simulations separately vary instrument relevance and IV invalidity. We also include GBDT and neural-network pricing benchmarks and estimate the sensitivity radius using Section D.3 in each simulation replication.

1. Synthetic data

Data-generating mechanism.

Each replication contains n=2,000n=2{,}000 observations. The covariates Xj∼Unif[−1,1]X_j\sim\mathrm{Unif}[-1,1], j=1,…,20j=1,\ldots,20, are mutually independent. Independently of X\boldsymbol X, we draw mutually independent U,Z,ξ∼Unif[−3,3]U,Z,\xi\sim\mathrm{Unif}[-\sqrt{3},\sqrt{3}] and V∼N(0,1)V\sim N(0,1). Let σ(t)=(1+e−t)−1\sigma(t)=(1+e^{-t})^{-1} and define

\begin{aligned} h(\boldsymbol x) &=0.8\sin(\pi x_1)+0.6x_2x_3+0.4(x_4^2-1/3)+0.4x_5+\frac{0.4}{\sqrt{12}}\sum_{j=9}^{20}x_j,\\\\ h_2(\boldsymbol x) &=0.8x_6+0.8x_7x_8+0.3x_9,\\\\ f_0(\boldsymbol x,p) &=80+4\tanh\\{h(\boldsymbol x)\\} -110(p-1)-2(p-1)^2+0.4\sin\\!\left\\{\frac{\pi(p-1)}{0.4}\right\\} \tanh\\{h_2(\boldsymbol x)\\}. \end{aligned}

Potential demand and historical observations are generated by

Y(p)=f0(X,p)+ϵ,ϵ=4U−ρ0Z+0.5ξ,P=0.8+0.4σ0.3h(X)+κZ+0.8U+0.6V,Y=Y(P),\begin{aligned} Y(p)&=f_0(\boldsymbol X,p)+\epsilon, \qquad \epsilon=4U-\rho_0 Z+0.5\xi,\\\\ P&=0.8+0.4\sigma\\{0.3h(\boldsymbol X)+\kappa Z+0.8U+0.6V\\}, \qquad Y=Y(P), \end{aligned}

where ρ0=(E![E[ϵ∣X,Z]2])1/2\rho_0= \left( \mathbb E\\!\left[ \\{\mathbb E[\epsilon\mid\boldsymbol X,Z]\\}^{2} \right]\right)^{1/2} and κ\kappa controls instrument relevance. Specifically, the historical pricing mechanism implies Corr⁡!(Z,log⁡P−0.81.2−P,|,X)=κ1+κ2.\operatorname{Corr}\\!\left( Z,\log\frac{P-0.8}{1.2-P} \\,\middle|\\,\boldsymbol X \right) =\frac{\kappa}{\sqrt{1+\kappa^2}}.

The proposed method uses the misspecified linear working demand space SK=span⁡1,x1,…,x20,p\mathcal{S}_K=\operatorname{span}\\{1,x_1,\ldots,x_{20},p\\}, which does not contain the nonlinear f0f_0. Furthermore, we set TL=span⁡1,x1,…,x20,z,z2,z3.\mathcal T_L=\operatorname{span} \\{1,x_1,\ldots,x_{20},z,z^2,z^3\\}.

Radius estimated using Section D.3 (Please see our response to Weakness 2).

We select the radius ρ\rho using demand range and slope restrictions that mimic practical domain knowledge (such as historical capacity limits and price sensitivity from past studies [6,7]). Specifically, we supply L=40,U=120,κ‾=106,κ‾=154L=40,\\ U=120,\\ \underline{\kappa}=106,\\ \overline{\kappa}=154 as prior bounds on E[Y(p)∣X=x,P=t]\mathbb E[Y(p)\mid\boldsymbol X=\boldsymbol x,P=t] and its derivative with respect to pp. These bounds are valid under the data-generating mechanism and provide partial information about demand without identifying f0f_0. In each replication, we estimate ρˉ0,cm{\bar\rho}_{0,\mathrm{cm}} and set ρ=ρˉ^0,cm\rho=\widehat{\bar\rho}_{0,\mathrm{cm}}.

Additional benchmarks and policy class.

We include GBDT [9] and NN pricing [10] (implemented via scikit-learn's GradientBoostingRegressor and MLPRegressor), which have been deployed in industrial pricing models at Zalando [11] and Meituan-Dianping [12], respectively. All methods use the same policy class \Pi= \left\\{ \pi_{\boldsymbol\theta}(\boldsymbol x) =1+\theta_0+\sum_{j=1}^{20}\theta_jx_j: \sum_{j=0}^{20}|\theta_j|\le0.2 \right\\}. The baseline policy is π0(x)=1\pi_0(\boldsymbol x)=1.

Experimental results.

Table 1 reports mean true revenue improvements across 30 replications, with policy values evaluated using 50,000 test covariate vectors. The mean estimated radius exceeds ρ0\rho_0 in every setting. Our method achieves the highest mean true revenue improvement across all settings.

Table 1. Mean true revenue improvement with standard errors in parentheses, across 30 training replications.

Panel A: Instrument relevance

Method(κ=0.45\kappa=0.45, ρ0=3\rho_0=3)(κ=0.15\kappa=0.15, ρ0=3\rho_0=3)(κ=0.05\kappa=0.05, ρ0=3\rho_0=3)
Proposed (estimated radius)1.933 (0.0009)1.920 (0.0010)1.914 (0.0011)
OLS1.242 (0.0568)-1.878 (0.0616)-2.496 (0.0686)
TLS-IV-0.463 (0.8336)-0.115 (0.6248)-3.070 (0.9271)
K-IV1.862 (0.0833)-5.539 (0.9670)-6.181 (1.0182)
DF-IV-10.495 (0.0004)-10.495 (0.0004)-10.495 (0.0004)
GBDT pricing0.168 (0.1705)-7.346 (0.7273)-9.973 (0.3702)
NN pricing0.234 (0.1221)-1.058 (0.1337)-1.796 (0.1382)
Mean estimated radius ρˉ^0,cm\widehat{\bar\rho}_{0,\mathrm{cm}}4.60394.89345.0266

Panel B: Instrument invalidity

Method(κ=0.45\kappa=0.45, ρ0=0\rho_0=0)(κ=0.45\kappa=0.45, ρ0=1.5\rho_0=1.5)(κ=0.45\kappa=0.45, ρ0=3\rho_0=3)
Proposed (estimated radius)1.935 (0.0009)1.946 (0.0006)1.933 (0.0009)
OLS-1.104 (0.0679)0.001 (0.0745)1.242 (0.0568)
TLS-IV0.136 (0.6790)-0.068 (0.7599)-0.463 (0.8336)
K-IV-2.142 (0.5886)1.776 (0.0243)1.862 (0.0833)
DF-IV-10.495 (0.0004)-10.495 (0.0004)-10.495 (0.0004)
GBDT pricing-9.475 (0.5707)-0.209 (0.1600)0.168 (0.1705)
NN pricing-0.778 (0.0841)-0.172 (0.0912)0.234 (0.1221)
Mean estimated radius ρˉ^0,cm\widehat{\bar\rho}_{0,\mathrm{cm}}3.39553.39564.6039

2. Real data

We use ready-to-eat breakfast cereal sales data from the IRI Infoscan database [13]. The sample contains 990 observations for 25 cereal brands in the Boston and San Francisco metropolitan markets over 20 quarters. Each observation summarizes one brand's sales in one market during one quarter. We use P=priceP=\texttt{price} and recover quantity demand by dividing sales revenue by price, Y=sales/priceY=\texttt{sales}/\texttt{price}. After dummy coding, the covariate vector X∈R30\boldsymbol X\in\mathbb R^{30} contains 24 brand indicators, one city indicator, three seasonal indicators, and two continuous covariates for quarterly advertising expenditure and the average price of other cereal brands in the same market and quarter.

We use two instruments based on the same brand's same-quarter prices in other markets: (i) the average price across New England markets outside Boston (for Boston) and across northern California markets outside San Francisco (for San Francisco); and (ii) the price in the other study market (San Francisco for Boston, and Boston for San Francisco). These instruments are relevant because the same brand shares common production costs across markets. However, as noted by Nevo and Rosen [13], regional or national promotions and shifts in consumer preferences can correlate other-market prices with unobserved local demand (ϵ\epsilon), violating instrument exogeneity.

Experimental setting.

We split the observations chronologically into 590 training observations and 400 test observations. The training data are used to fit the benchmark demand models, construct the empirical ambiguity set, and learn pricing policies. Any hyperparameter selection uses training data only. The learned policies are held fixed during testing. The test data are used to construct the evaluation ambiguity set and compute empirical worst-case revenue improvements. All methods use the same linear pricing policy class, allowing prices to differ across brands, cities, and seasons. The baseline price π0(x)\pi_0(\boldsymbol x) is the training median price for the corresponding brand and city. Recommended prices satisfy 0.9π0(x)≤π(x)≤1.1π0(x),0.9\pi_0(\boldsymbol x) \le \pi(\boldsymbol x) \le 1.1\pi_0(\boldsymbol x), and must also remain within that brand's observed training price range in that city.

Evaluation.

In this real-data application, the true demand function f0f_0 and counterfactual revenues are unknown. With invalid instruments, we cannot identify f0f_0 from the observed data. We therefore compare policies on the test set by their worst-case revenue improvement over the baseline:

\min_{f\in\mathcal F(\rho;\mathcal S_K,\mathcal T_L)} \left\\{ \widehat V_f(\widehat\pi) - \widehat V_f(\pi_0) \right\\},

where the sieve spaces satisfy SK=span⁡1,x1,…,x30,p,p2\mathcal S_K=\operatorname{span} \\{1,x_1,\ldots,x_{30},p,p^2\\} and TL=span⁡1,x1,…,x30,z1,z2,z12,z22,z1z2\mathcal T_L=\operatorname{span} \\{1,x_1,\ldots,x_{30},z_1,z_2,z_1^2,z_2^2,z_1z_2\\}.

Table 2. Empirical worst-case revenue improvement on the held-out IRI Infoscan breakfast cereal sales data. The reported ρ\rho is the sensitivity radius normalized by the training-sample standard deviation of demand.

Methodρ=0.1\rho=0.1ρ=0.2\rho=0.2ρ=0.4\rho=0.4ρ=0.8\rho=0.8
Proposed18,18014,9039,6012,494
OLS−37,166-37,166−110,893-110,893−255,991-255,991−545,050-545,050
TLS-IV−4,947-4,947−46,951-46,951−129,618-129,618−294,304-294,304
K-IV−12,120-12,120−31,291-31,291−69,020-69,020−144,184-144,184
DF-IV−86,293-86,293−176,674-176,674−354,549-354,549−708,905-708,905
GBDT pricing−31,250-31,250−87,219-87,219−197,370-197,370−416,807-416,807
NN pricing−26,955-26,955−72,320-72,320−161,599-161,599−339,459-339,459

Results.

As shown in Table 2, the proposed method achieves the largest empirical worst-case revenue improvement at each reported radius and remains positive throughout the displayed range. Its improvement decreases from 18,180 at ρ=0.1\rho=0.1 to 2,494 at ρ=0.8\rho=0.8. These results illustrate how the proposed method retains positive worst-case improvement as the specified uncertainty increases in this real-data application.

Reviewer_zrJhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies offline pricing when IVs are invalid. It relaxes the invalidity assumption by constructing an ambiguity set and looks for a robust policy that optimises its worst case scenario. The authors give an optimisation algorithm to achieve this and accompany it with theoretical results. The approach is validated on synthetic datasets.

Critical strengths

The main merit of the paper is the problem the authors are trying to solve. Training robust policies when instrument variable are invalid is aligned with realistic scenarios and can help practitioners mitigate deployment risks.

Critical weaknesses

Even if i'm very familiar with the literature, I found the paper near impossible to read, and in this era of AI assisted papers, readability is becoming one of the main differentiators. The authors obfuscate the results into very complex formulations and do a poor job bringing the core insights forward. The propositions can sometimes be huge paragraphs of a raw dump of cumbersome theoretical results. For example, no human will look at Theorem 3 and 4 and finds them elegant and well presented.

Questions for authors

Correct me if i'm wrong, but even after all the machinery introduced, we still have no clear way of fixing the ambiguity set radius ρ\rho as you need knowledge of the unknown ρ0\rho_0? This makes the proposed approach impractical.

7 comments

Response to Weakness 1:

We sincerely thank the reviewer for this constructive feedback. We will substantially revise the manuscript to improve readability and bring the core insights to the forefront. Specifically, we will add descriptive titles and clear intuitive explanations to our theoretical results and decompose the error bounds into interpretable components.

First, we will revise Proposition 1 as follows:

To illustrate why approximately valid instruments can still be informative, we first consider the case f0∈SKf_0\in\mathcal S_K.

Proposition 1 (When an imperfect instrument remains informative).

Suppose that f0∈SKf_0\in\mathcal S_K, E[Y2]<∞\mathbb E[Y^2] < \infty, and QKQ_K and GKG_K are positive definite. Let λmin⁡\lambda_{\min} denote the smallest eigenvalue of QK−1/2GKQK−1/2Q_K^{-1/2}G_KQ_K^{-1/2}. If \sqrt{\mathbb{E}\left[ \left\\{ \mathbb{E}(\epsilon\mid\boldsymbol X,Z) \right\\}^2 \right]} < \rho < \frac{\sqrt{\lambda_{\min}}}{2} \sqrt{\mathbb{E}\left[ \left\\{\mathbb{E}(\epsilon\mid\boldsymbol X,P)\right\\}^2 \right]}, then f0∈F(ρ;SK)f_0\in\mathcal F(\rho;\mathcal S_K), and every f∈F(ρ;SK)f\in\mathcal F(\rho;\mathcal S_K) satisfies

E![f(X,P)−f0(X,P)2]<E![fols(X,P)−f0(X,P)2].\mathbb E\\!\left[ \\{f(\boldsymbol X,P)-f_0(\boldsymbol X,P)\\}^2 \right] < \mathbb E\\!\left[ \\{f_{\mathrm{ols}}(\boldsymbol X,P) -f_0(\boldsymbol X,P)\\}^2 \right].

Proposition 1 demonstrates that every demand function in the ambiguity set achieves a smaller mean squared error than folsf_{\mathrm{ols}}. The lower bound on ρ\rho ensures that the set contains the true demand function, while the upper bound accounts for instrument strength through λmin⁡\lambda_{\min}. Thus, an imperfect instrument remains informative whenever its exogeneity violation E[E(ϵ∣X,Z)2]\sqrt{\mathbb{E}[\\{\mathbb{E}(\epsilon\mid\boldsymbol X,Z)\\}^2]} is sufficiently small relative to the endogeneity bias E[E(ϵ∣X,P)2]\sqrt{\mathbb{E}[\\{\mathbb{E}(\epsilon\mid\boldsymbol X,P)\\}^2]}.

Second, we will revise Theorems 3 and 4 to make their structure and main results much easier to follow:

We next show that the estimated robust policy achieves nearly the largest worst-case revenue improvement among all policies in Π\Pi.

Theorem 3 (Vanishing worst-case regret).

Suppose that Assumptions 2 and 3 hold (Assumption 3 is stated in Appendix B.6). Define

E(n,K,L)=τK,L2ζK,L2log⁡nn⏟statistical error+τK,LζK,LbL2⏟variational approximation error.\mathcal E(n,K,L) = \underbrace{ \tau_{K,L}^2\zeta_{K,L}^2 \sqrt{\frac{\log n}{n}} }_{\text{statistical error}} + \underbrace{ \tau_{K,L}\zeta_{K,L}b_{L}^2 }_{\text{variational approximation error}}.

Then

Reg⁡mmπ^K,L⋆(ρ);ρ=OpE(n,K,L).\operatorname{Reg}_{\mathrm{mm}} \\{\widehat\pi_{K,L}^{\star}(\rho);\rho\\} = O_p\\{\mathcal E(n,K,L)\\}.

The bound E(n,K,L)\mathcal E(n,K,L) captures the trade-off in choosing the sieve dimensions (K,L)(K,L): a larger LL reduces the variational approximation error bL2b_{L}^2, whereas larger KK and LL increase finite-sample statistical error through the factors τK,L\tau_{K,L} and ζK,L\zeta_{K,L}. Under Assumption 2, E(n,K,L)→0\mathcal E(n,K,L)\to 0. Thus, the worst-case performance gap between π^K,L⋆(ρ)\widehat\pi_{K,L}^{\star}(\rho) and the best policy in Π\Pi over F(ρ;SK)\mathcal F(\rho;\mathcal S_K) vanishes in probability.

While Theorem 3 evaluates worst-case performance within the demand class SK\mathcal S_K, Theorem 4 establishes a guaranteed revenue improvement under the true demand function f0f_0.

Theorem 4 (Target improvement under the true demand function).

Suppose that Assumptions 2 and 3 hold, and assume additionally that:

(i) Demand approximation error. The pair (X,P)(\boldsymbol X,P) is supported on a compact rectangular domain Df⊂Rdf\mathcal D_f\subset\mathbb R^{d_f}, and f0∈Hsf(Df,Bf)f_0\in\mathcal H^{s_f}(\mathcal D_f,B_f), the Hölder class with smoothness sfs_f and radius BfB_f (see Appendix B.5). The sieve approximation error satisfies

inf⁡f~∈SK∣f~−f0∣∞≤αK,αK⟶0,τK,LζK,LαK⟶0.\inf_{\widetilde f\in\mathcal S_K} \\|\widetilde f-f_0\\|_\infty\le\alpha_K, \qquad \alpha_K\longrightarrow0, \qquad \tau_{K,L}\zeta_{K,L}\alpha_K\longrightarrow0.

(ii) Instrument exogeneity violation. The actual instrument exogeneity violation is bounded by the population gain radius:

\left\\| \mathbb E[\epsilon\mid\boldsymbol X,Z] \right\\|_2 \le\rho_{\mathrm{gain},K}(\delta).

Then

\begin{aligned} &V_{f_0}\\!\left[ \widehat\pi_{K,L}^{\star} \\{\widehat\rho_{\mathrm{gain},K,L}(\delta)\\} \right]-V_{f_0}(\pi_0) \\\\ &\qquad\ge \delta- O_p\\!\left\\{ \mathcal E(n,K,L) +\underbrace{\tau_{K,L}\zeta_{K,L}\alpha_K}_{\text{demand approximation error}} \right\\}. \end{aligned}

Here, αK\alpha_K bounds the error from approximating the true demand function f0f_0 by a function in SK\mathcal S_K. When f0∈SKf_0\in\mathcal S_K, we can take αK=0\alpha_K=0, so the additional term τK,LζK,LαK\tau_{K,L}\zeta_{K,L}\alpha_K disappears. Theorem 4 shows that the learned policy achieves an expected revenue improvement of at least δ−op(1)\delta-o_p(1) over the baseline under the true demand function.

Response to Question 1:

We thank the reviewer for this important comment. We have developed a new extension (Section D) showing that specifying the ambiguity set radius ρ\rho does not require knowing the exact value of ρ0\rho_0; instead, we can derive an estimable upper bound ρˉ0\bar{\rho}_0 on ρ0\rho_0 by combining the observational distribution with shape restrictions on potential demand. Such restrictions can be informed by structural demand models [2], known market-size information [5], and expert knowledge [8]. Moreover, relevant pricing experiments [6] and external empirical studies [7] can inform restrictions on mean price responses. Consequently, practitioners have a data-driven way to choose the radius by setting ρ=ρˉ0\rho = \bar{\rho}_0, which directly guarantees ρ0≤ρ\rho_0\le\rho.

We will add the following extension to the revised manuscript to show how to construct an upper bound ρˉ0\bar{\rho}_0 on ρ0\rho_0 in practice.

D Extension: Bounding Instrument Invalidity Using Demand Restrictions

The pricing literature commonly imposes restrictions on demand levels and price sensitivity to characterize the unknown demand function, including boundedness, monotonicity, and slope restrictions [1,2]. Motivated by this approach, we show how domain knowledge about potential demand or its conditional mean can be combined with the observational distribution to derive simple and estimable upper bounds on the actual instrument invalidity, ρ0=(E![E[ϵ∣X,Z]2])1/2.\rho_0 = \left( \mathbb E\\!\left[ \\{\mathbb E[\epsilon\mid\boldsymbol X,Z]\\}^2 \right] \right)^{1/2}.

D.1 Bounded and Monotone Potential Demand

We first consider applications in which domain knowledge provides bounds on potential demand and supports a nonincreasing price response.

Suppose that the known finite constants L≤UL\le U satisfy, almost surely, L≤Y(p)≤UL\le Y(p)\le U for all admissible prices pp, and Y(p1)≥Y(p2)Y(p_1)\ge Y(p_2) for all admissible prices p1<p2p_1 < p_2. These assumptions are economically natural. For example, in unit-demand reservation-value models [3], a customer purchases at most one unit if and only if the price does not exceed their reservation value; thus, potential demand naturally satisfies Y(p)∈0,1Y(p)\in\\{0,1\\} (with L=0L=0 and U=1U=1) and is nonincreasing in price.

For any fixed price pp, consistency and monotonicity imply that Y(p)≥YY(p)\ge Y whenever P≥pP\ge p, and Y(p)≤YY(p)\le Y whenever P≤pP\le p. Then, we have

Y1P≥p+L1P<p≤Y(p)≤Y1P≤p+U1P>p.Y\mathbf{1}\\{P\ge p\\} + L\mathbf{1}\\{P < p\\} \le Y(p)\le Y\mathbf{1}\\{P\le p\\} + U\mathbf{1}\\{P > p\\}.

It follows that

E![Y1P≥p+L1P<p∣X=x]≤f0(x,p)≤E![Y1P≤p+U1P>p∣X=x].\begin{aligned} \mathbb{E}\\!\left[ Y\mathbf{1}\\{P\ge p\\} + L\mathbf{1}\\{P < p\\} \mid\boldsymbol X=\boldsymbol x \right] \le & f_0(\boldsymbol x,p) \\\\ \le & \mathbb{E}\\!\left[ Y\mathbf{1}\\{P\le p\\} + U\mathbf{1}\\{P > p\\} \mid\boldsymbol X=\boldsymbol x \right]. \end{aligned}

Let

L(x,p)=E![Y1P≥p+L1P<p∣X=x],U(x,p)=E![Y1P≤p+U1P>p∣X=x].\begin{aligned} \mathcal L(\boldsymbol x,p) &= \mathbb{E}\\!\left[ Y\mathbf{1}\\{P\ge p\\} + L\mathbf{1}\\{P < p\\} \mid\boldsymbol X=\boldsymbol x \right], \\\\ \mathcal U(\boldsymbol x,p) &= \mathbb{E}\\!\left[ Y\mathbf{1}\\{P\le p\\} + U\mathbf{1}\\{P > p\\} \mid\boldsymbol X=\boldsymbol x \right]. \end{aligned}

Then, we have L(X,P)≤f0(X,P)≤U(X,P).\mathcal L(\boldsymbol X,P) \le f_0(\boldsymbol X,P) \le\mathcal U(\boldsymbol X,P). Since ϵ=Y−f0(X,P)\epsilon=Y-f_0(\boldsymbol X,P), we have

Y−U(X,P)≤ϵ≤Y−L(X,P).Y-\mathcal U(\boldsymbol X,P) \le\epsilon\le Y-\mathcal L(\boldsymbol X,P).

It follows that

E![Y−U(X,P)∣X,Z]≤E[ϵ∣X,Z]≤E![Y−L(X,P)∣X,Z].\begin{aligned} \mathbb{E}\\!\left[ Y-\mathcal U(\boldsymbol X,P) \mid\boldsymbol X,Z \right] \le \mathbb{E}[\epsilon\mid\boldsymbol X,Z] \le \mathbb{E}\\!\left[ Y-\mathcal L(\boldsymbol X,P) \mid\boldsymbol X,Z \right]. \end{aligned}

Therefore, we obtain the following upper bound on ρ02\rho_0^2:

\begin{aligned} \rho_0^2 &= \mathbb{E}\left[ \\{\mathbb{E}[\epsilon\mid\boldsymbol X,Z]\\}^2 \right] \\\\ &\le \mathbb{E}\Bigg[ \max\Bigg\\{ \left( \mathbb{E}\\!\left[ Y-\mathcal U(\boldsymbol X,P) \mid\boldsymbol X,Z \right] \right)^2, \left( \mathbb{E}\\!\left[ Y-\mathcal L(\boldsymbol X,P) \mid\boldsymbol X,Z \right] \right)^2 \Bigg\\} \Bigg] \\\\ &= \bar{\rho}_0^2. \end{aligned}

Hence, ρ0≤ρˉ0\rho_0\le\bar{\rho}_0. The upper bound ρˉ02\bar{\rho}_0^2 depends only on the observational distribution and the known demand bounds LL and UU, making it directly estimable from data.

Reviewer_hGVROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies offline pricing when historical prices are endogenous and instrumental variables may be invalid. It constructs a set of demand functions that allows bounded violations of the IV condition and chooses a policy that maximizes worst-case revenue improvement over a baseline.

Using basis approximations, the authors derive a closed-form solution for the inner minimization. They also find the largest violation radius that supports a given improvement target and establish convergence guarantees. Synthetic experiments compare the proposed method with OLS and several IV-based methods.

Critical strengths
  • Instrument validity can be difficult to justify in practice, and accounting for possible violations is useful for pricing decisions. The focus on improving an existing policy is clear contribution.

  • The closed-form expression separates estimated revenue improvement from a robustness penalty, making the objective easier to evaluate. The theoretical analysis is quite long, covering several results including estimation error, basis approximation, and weak identification.

Critical weaknesses
  • There are a few inconsistencies. To begin with, the paper starts with a bounded Holder class but later uses basis expansions with unrestricted coefficients. Adding more basis functions can therefore introduce increasingly extreme demand functions into the ambiguity set, potentially reducing the worst-case improvement to zero.
    • Theorem 3 shows that the learned policy approaches the best robust policy within the chosen demand class, but this guarantee is less informative if that best policy simply retains the baseline.
    • Assumptions 2(i) and 3(ii) are thus imposed but not sufficiently justified or explained when they hold.
  • Numerical results do not deliver very convincing evidence. The selected radius measures how much IV violation a target improvement can tolerate. But does not establish that the actual violation is below that level, which remains an assumption in Theorem 4. In particular, all four benchmarks optimize nominal demand estimates, so their poor performance under invalid instruments is largely expected. The main and airline experiments also use small, correctly specified demand models.
Questions for authors

Please see weaknesses.

In practice, how to judge whether the actual IV error is below the chosen radius when that error is unknown?

13 comments

We sincerely thank the reviewer for the insightful comments and constructive feedback. We have carefully addressed all of your concerns and provide our point-by-point responses below.

Q1: There are a few inconsistencies. To begin with, the paper starts with a bounded Holder class but later uses basis expansions with unrestricted coefficients.

Response to Q1:

We thank the reviewer for raising this important point. Although we use basis expansions with unrestricted coefficients, the coefficients of the demand functions in the ambiguity set F(ρ;SK,TL)\mathcal F(\rho;\mathcal S_K,\mathcal T_L) are bounded for each fixed pair (K,L)(K,L).

To see this clearly, we write out the double-sieve ambiguity set:

\mathcal F(\rho;\mathcal S_K,\mathcal T_L) = \left\\{f_{\boldsymbol\beta}\in\mathcal S_K:\max_{g\in\mathcal T_L}\left[2\mathbb E\\!\left[g(\boldsymbol X,Z)\\{Y-f_{\boldsymbol\beta}(\boldsymbol X,P)\\}\right]-\mathbb E\\{g(\boldsymbol X,Z)^2\\}\right]\le\rho^2\right\\}.

By solving the inner quadratic maximization over the finite-dimensional space TL\mathcal T_L, we obtain the following equivalent representation of the set:

\mathcal F(\rho;\mathcal S_K,\mathcal T_L) = \left\\{f_{\boldsymbol\beta}\in\mathcal S_K:(\boldsymbol\beta-\overline{\boldsymbol\beta}_{K,L})^\top G_{K,L}(\boldsymbol\beta-\overline{\boldsymbol\beta}_{K,L})\le\rho^2-q_{K,L}\right\\},

where

qK,L=E[ψLY−ϕK⊤βˉK,L]⊤E[ψLψL⊤]−1E[ψLY−ϕK⊤βˉK,L],q_{K,L} = \mathbb E\left[\boldsymbol\psi_L\\{Y-\boldsymbol\phi_K^\top\bar{\boldsymbol\beta}_{K,L}\\}\right]^\top\mathbb E[\boldsymbol\psi_L\boldsymbol\psi_L^\top]^{-1}\mathbb E\left[\boldsymbol\psi_L\\{Y-\boldsymbol\phi_K^\top\bar{\boldsymbol\beta}_{K,L}\\}\right],

GK,L=E[ϕKψL⊤]E[ψLψL⊤]−1E[ψLϕK⊤],G_{K,L} = \mathbb E[\boldsymbol\phi_K\boldsymbol\psi_L^\top]\mathbb E[\boldsymbol\psi_L\boldsymbol\psi_L^\top]^{-1}\mathbb E[\boldsymbol\psi_L\boldsymbol\phi_K^\top],

and ψL\boldsymbol\psi_L and ϕK\boldsymbol\phi_K are vectors of basis functions.

Consequently, the set of feasible coefficients β\boldsymbol\beta is bounded within a compact ellipsoid centered at β‾K,L\overline{\boldsymbol\beta}_{K,L} with its radius governed by ρ2−qK,L\rho^2-q_{K,L}.

Q2: Adding more basis functions can therefore introduce increasingly extreme demand functions into the ambiguity set, potentially reducing the worst-case improvement to zero.

Response to Q2:

We consider two approaches to addressing this concern. A simple approach is to control the magnitude of feasible coefficients through quadratic regularization. Specifically, we can replace F(ρ;SK,TL)\mathcal F(\rho;\mathcal S_K,\mathcal T_L) with

\mathcal F_{\eta}(\rho;\mathcal S_K,\mathcal T_L) = \left\\{ f_{\boldsymbol\beta}\in\mathcal S_K: (\boldsymbol\beta-\overline{\boldsymbol\beta}_{K,L})^\top G_{K,L} (\boldsymbol\beta-\overline{\boldsymbol\beta}_{K,L}) +\eta\\|\boldsymbol\beta\\|_2^2 \le \rho^2-q_{K,L} \right\\}, \qquad \eta > 0.

Every feasible coefficient β\boldsymbol{\beta} in Fη(ρ;SK,TL)\mathcal F_{\eta}(\rho;\mathcal S_K,\mathcal T_L) satisfies ∣β∣2≤ρ/η\\|\boldsymbol\beta\\|_2\le\rho/\sqrt{\eta}. The additional quadratic term preserves the ellipsoidal structure and allows the same closed-form solution approach proposed in the paper.

Alternatively, when domain knowledge provides bounds on the true demand function, we can incorporate these bounds through coefficient restrictions.

Section 5: Incorporating domain knowledge on demand bounds

In some applications, domain knowledge indicates that demand is nonnegative and bounded above by a known constant M>0M > 0, such as a market-size limit. To accommodate this information, we consider the following ambiguity set:

\mathcal F^{\rm box}_{K,L}(\rho,M) = \left\\{ f_{\boldsymbol\beta}\in\mathcal S_K: \begin{array}{l} (\boldsymbol\beta-\bar{\boldsymbol\beta}_{K,L})^\top G_{K,L} (\boldsymbol\beta-\bar{\boldsymbol\beta}_{K,L}) \le \rho^2-q_{K,L},\\\\[2pt] \boldsymbol{0}\le\boldsymbol\beta\le M\boldsymbol{1} \end{array} \right\\}.

Here, G_{K,L} = \mathbb E[\boldsymbol\phi_K\boldsymbol\psi_L^\top] \left\\{ \mathbb E[\boldsymbol\psi_L\boldsymbol\psi_L^\top] \right\\}^{-1} \mathbb E[\boldsymbol\psi_L\boldsymbol\phi_K^\top], and the basis satisfies ϕK(x,p)≥0\boldsymbol\phi_K(\boldsymbol x,p)\ge0 and 1⊤ϕK(x,p)=1\boldsymbol{1}^\top\boldsymbol\phi_K(\boldsymbol x,p)=1 (e.g., unnormalized B-spline bases, Bernstein bases, partition indicator bases, and their tensor products). Consequently, every demand function in the ambiguity set FK,Lbox(ρ,M)\mathcal F^{\rm box}_{K,L}(\rho,M) satisfies

0≤fβ(x,p)=∑j=1Kβjϕj,K(x,p)≤M∑j=1Kϕj,K(x,p)=M.0 \le f_{\boldsymbol\beta}(\boldsymbol x,p) = \sum_{j=1}^{K}\beta_j\phi_{j,K}(\boldsymbol x,p) \le M\sum_{j=1}^{K}\phi_{j,K}(\boldsymbol x,p) = M.

We then consider the following robust policy optimization problem

\max_{\pi\in\Pi} \min_{f_{\boldsymbol\beta}\in \mathcal F^{\rm box}_{K,L}(\rho,M)} \left\\{ V_{f_{\boldsymbol\beta}}(\pi) - V_{f_{\boldsymbol\beta}}(\pi_0) \right\\}. \tag{Q2.1}

The following corollary provides an exact dual reformulation.

Corollary 1. Suppose that E[ψLψL⊤]\mathbb{E}[\boldsymbol\psi_L\boldsymbol\psi_L^\top] is positive definite and E[ψLϕK⊤]\mathbb{E}[\boldsymbol\psi_L\boldsymbol\phi_K^\top] has full column rank. Assume that there exists a coefficient vector β\boldsymbol\beta satisfying

0<β<M1,(β−βˉK,L)⊤GK,L(β−βˉK,L)<ρ2−qK,L.\boldsymbol{0} < \boldsymbol\beta < M\boldsymbol{1}, \qquad (\boldsymbol\beta-\bar{\boldsymbol\beta}_{K,L})^\top G_{K,L} (\boldsymbol\beta-\bar{\boldsymbol\beta}_{K,L}) < \rho^2-q_{K,L}.

Then, for every policy π∈Π\pi\in\Pi,

\begin{aligned} &\min_{f_{\boldsymbol\beta}\in \mathcal F^{\rm box}_{K,L}(\rho,M)} \left\\{ V_{f_{\boldsymbol\beta}}(\pi) - V_{f_{\boldsymbol\beta}}(\pi_0) \right\\} \\\\ &= \max_{\boldsymbol\lambda\in\mathbb R^K} \Bigg\\{ V_{f_{\bar{\boldsymbol\beta}_{K,L}}}(\pi) - V_{f_{\bar{\boldsymbol\beta}_{K,L}}}(\pi_0) + \boldsymbol\lambda^\top\bar{\boldsymbol\beta}_{K,L} \\\\ &\qquad - \sqrt{\rho^2-q_{K,L}}\\, \left\\| G_{K,L}^{-1/2} \left[ \nabla_{\boldsymbol\beta} \left\\{ V_{f_{\boldsymbol\beta}}(\pi) - V_{f_{\boldsymbol\beta}}(\pi_0) \right\\} + \boldsymbol\lambda \right] \right\\|_2 \\\\ &\qquad - M\sum_{j=1}^{K}\max\\{\lambda_j,0\\} \Bigg\\}. \end{aligned}

Therefore, the robust policy optimization problem (Q2.1) can be solved by

\begin{aligned} \max_{\pi,\boldsymbol\lambda}\quad & V_{f_{\bar{\boldsymbol\beta}_{K,L}}}(\pi) - V_{f_{\bar{\boldsymbol\beta}_{K,L}}}(\pi_0) + \boldsymbol\lambda^\top\bar{\boldsymbol\beta}_{K,L} \\\\ &- \sqrt{\rho^2-q_{K,L}}\\, \left\\| G_{K,L}^{-1/2} \left[ \nabla_{\boldsymbol\beta} \left\\{ V_{f_{\boldsymbol\beta}}(\pi) - V_{f_{\boldsymbol\beta}}(\pi_0) \right\\} + \boldsymbol\lambda \right] \right\\|_2 \\\\ &- M\sum_{j=1}^{K}\max\\{\lambda_j,0\\} \\\\ \text{subject to}\quad & \pi\in\Pi, \qquad \boldsymbol\lambda\in\mathbb R^K. \end{aligned}

Q3: Theorem 3 shows that the learned policy approaches the best robust policy within the chosen demand class, but this guarantee is less informative if that best policy simply retains the baseline.

Response to Q3:

We thank the reviewer for this comment. To guarantee a strictly positive gain, Theorem 4 establishes a true policy value improvement of at least δ−op(1)\delta-o_p(1) for a user-specified target δ>0\delta > 0, provided that

E![E(ϵ∣X,Z)2]≤ρgain,K(δ)2\mathbb E\\!\left[ \\{\mathbb E(\epsilon\mid\boldsymbol X,Z)\\}^2 \right] \le \rho_{\mathrm{gain},K}(\delta)^2

and the remaining regularity conditions hold.

In practice, we can derive an upper bound ρˉ0\bar{\rho}_0 on the actual IV error ρ0\rho_0 by combining the observational distribution with shape restrictions on potential demand. Such restrictions can be informed by structural demand models [2], outcome definitions and unit-demand purchase models [3], known market-size information [5], and expert knowledge [8]. Moreover, relevant pricing experiments [6] and external empirical studies [7] can inform restrictions on mean price responses. Consequently, we can judge whether the actual IV error is below the chosen radius ρ\rho simply by checking if ρˉ0≤ρ\bar{\rho}_0\le\rho (which guarantees ρ0≤ρ\rho_0\le\rho). If the available evidence is insufficient to support the condition ρˉ0≤ρgain,K(δ)\bar{\rho}_0 \le\rho_{\mathrm{gain},K}(\delta), the desired target revenue improvement δ\delta cannot be certified. This remains highly informative for decision-making: it warns the decision-maker that the current data and instrument quality cannot justify the initial desired revenue improvement. In response, decision-makers can adopt a revised target δ′\delta^{\prime} such that ρgain,K(δ′)=ρˉ0\rho_{\mathrm{gain},K}(\delta^{\prime})=\bar{\rho}_0 to obtain a robust policy.

We will add the following extension to the revised manuscript to show how to construct an upper bound ρˉ0\bar{\rho}_0 on ρ0\rho_0 in practice.

D Extension: Bounding Instrument Invalidity Using Demand Restrictions

The pricing literature commonly imposes restrictions on demand levels and price sensitivity to characterize the unknown demand function, including boundedness, monotonicity, and slope restrictions [1,2]. Motivated by this approach, we show how domain knowledge about potential demand or its conditional mean can be combined with the observational distribution to derive simple and estimable upper bounds on the actual instrument invalidity, ρ0=(E![E[ϵ∣X,Z]2])1/2.\rho_0 = \left( \mathbb E\\!\left[ \\{\mathbb E[\epsilon\mid\boldsymbol X,Z]\\}^2 \right] \right)^{1/2}.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.