acceptodds
Under review as a conference paper at ICLR 2027

Learning Advertiser Values from Strategic Bids in Platform Position Auctions

Abstract

Digital platforms allocate scarce user attention through position auctions, yet the advertiser values needed for platform design are not directly observed. In generalized second-price (GSP) auctions, bids are strategic and therefore cannot be interpreted as values. This poses a structural estimation problem: how can a platform recover latent advertiser value distributions from observed bids when analytical bid-to-value inversion is mechanism-specific or altogether unavailable? We propose a simulation-based estimator for non-truthful platform auctions. The method represents advertiser value distributions with flexible Bernstein-polynomial CDFs, computes approximate equilibrium bidding strategies for candidate primitives, and selects the primitives whose simulated bids match the observed bids in Wasserstein distance. A Bayesian-optimization-assisted SMC-ABC procedure guides the search while limiting the number of costly equilibrium computations. We validate the approach on benchmark auction simulations and on GSP laboratory data, and use the recovered value distributions to evaluate counterfactual platform-design choices such as optimal reserve prices. The framework offers a practical route to data-driven market design in digital platform auctions where standard inversion-based estimators are difficult to apply.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
platform auctions, generalized second-price auctions, structural estimation, sponsored search, market design, digital platforms, approximate bayesian computation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_q2dgOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper estimates advertiser value distributions from strategic GSP bids using Bernstein CDFs, a SODA equilibrium simulator, Wasserstein matching, and BO-assisted SMC-ABC. Its central motivation is portability to auction mechanisms where analytical inversion is difficult or unavailable.

Critical strengths

1.Well-motivated problem and appropriately modest positioning against specialized estimators.

2.Explicit discussion of approximation errors and assumptions required for convergence.

3.Lower average distributional error than KimPal in the reported uniform/beta simulations.

Critical weaknesses

1.Insufficient literature positioning and an overgeneralized identification claim. BOLFI [1] and the related auction-estimation work by Ewert & Bichler [2] are not discussed. Positioning the proposed framework relative to these works would help clarify its incremental contribution. Additionally, the introduction’s claim that generalized or unobserved scoring rules make primitives “fundamentally unidentified” needs qualification: Laffont et al. (2020), Econometrics of Scoring Auctions, provides identification results under stated assumptions. The paper should distinguish the absence of a tractable analytical inversion from non-identification. A simulation-based estimator does not by itself resolve the latter, and Section 4 appropriately treats identification as a maintained condition.

2.The central portability claim lacks direct experimental validation. The experiments vary distributions and auction parameters within GSP settings, but do not demonstrate recovery under a mechanism explicitly identified as lacking an available analytical inversion. The comparisons establish performance in benchmark environments, rather than the claimed advantage when mechanism-specific inversion is unavailable. A recovery experiment under such a mechanism would directly support the paper’s central motivation; outperforming KimPal in its applicable settings is not necessary.

3.Lab-model compatibility remains insufficiently documented. Bae & Kagel’s dataset includes both static complete-information and dynamic incomplete-information auctions, with valuation draws subject to ratio constraints. Section 5.2 does not specify which information treatments and bidding rounds enter estimation, or how the constrained value-generation process is reconciled with the model’s independence assumption. The information structure and within-auction feedback also require justification for interpreting these observations through a static symmetric BNE model. Although Section 3 acknowledges possible equilibrium nonexistence and states that recovered strategies satisfy the maintained restriction, treatment-specific evidence supporting equilibrium validity is not reported. These gaps limit the interpretation of the lab results as structural recovery, rather than distributional fit under a potentially misspecified behavioral model.

[1] Michael U. Gutmann and Jukka Corander. Bayesian Optimization for Likelihood-Free Inference of Simulator-Based Statistical Models. JMLR, 17(125):1–47, 2016. Official record [2] Markus Ewert and Martin Bichler. Structural Estimation of Auction Data through Equilibrium Learning and Optimal Transport. WI 2025 conference contribution. Official conference program

Questions for authors

See the weakness

1 comment

Thank you for your constructive review.

W1a: Related work. BOLFI (Gutmann & Corander, 2016) uses a Gaussian-process surrogate to guide simulations and approximate the likelihood-free posterior. BOSMC instead uses a surrogate to reject selected SMC-ABC proposals before costly equilibrium simulation; proposals passing this screening are evaluated using their simulated rather than surrogate-predicted discrepancies. Importantly, this screening changes the effective proposal distribution. Without accounting for it in the weighting scheme, the resulting particles need not target the same ABC posterior as standard SMC-ABC. We therefore do not claim that BOSMC preserves the exact SMC-ABC guarantees. Screening trades computational savings against possible approximation bias. We will clarify the screening rule and its limitations.

Ewert & Bichler (2025) also combine equilibrium learning and optimal transport with a GPV-style estimation framework. The distinction is not the use of equilibrium computation or Wasserstein distance per se, but the role of inversion: their method refines a two-step inversion-based estimator, whereas ours matches bids generated by forward equilibrium simulation without implementing a mechanism-specific inverse mapping. We will clarify this relationship.

W1b: Identification vs. inversion. We agree that our introduction conflates these issues. Laffont et al. (2020) establish identification under specified assumptions; our text should not suggest that the absence of a tractable inversion formula implies non-identification. Simulation does not create identification, but can exploit it without explicitly implementing an inverse.

For our symmetric IPV GSP benchmark with known CTRs, linear scoring, and a strictly increasing symmetric BNE, identification follows from the equilibrium first-order condition under suitable regularity assumptions. It yields v=ξ(b;G)v=\xi(b;G), where GG is the bid distribution (Kim & Pal, 2025). If F1F_1 and F2F_2 generated the same GG under these restrictions, then F1(ξ(b;G))=G(b)=F2(ξ(b;G))F_1(\xi(b;G))=G(b)=F_2(\xi(b;G)) on the relevant support, implying F1=F2F_1=F_2. With fixed Bernstein degree and support, an injective coefficient representation likewise identifies θ\theta. We will state these assumptions explicitly. A platform knows its own scoring rule, but that knowledge alone does not ensure identification in more general mechanisms; identification must be assessed separately from tractability.

W2: Portability. We agree that an experiment on a mechanism without an available analytical inversion would most directly validate the motivating application, and that our current experiments do not establish this empirically. Instead, they demonstrate recovery in benchmark environments where comparison with a specialized estimator is possible. Portability is an architectural feature: the forward-simulation estimation procedure can be retained while adapting the equilibrium simulator and any mechanism-specific model components, subject to identification and suitable equilibrium computation. We will narrow the corresponding claims.

W3: Laboratory data. Thank you for identifying missing details. We use the dynamic incomplete-information treatments with CTRs 11-10 and 11-3, each comprising three groups. Within each group, we pool bidding rounds (960–1,920 value–bid observations per group, 8,624 in total), estimate a symmetric three-bidder model separately, and report treatment-level means and standard deviations across groups in Table 3. We will document these choices in Section 5.2.

The ratio restriction in Bae & Kagel (2019) applies to the static complete-information auctions, not the dynamic incomplete-information treatments we use. In the latter, values are independently drawn from [1,100][1,100], consistent with the IPV specification. We agree, however, that repeated bidding and within-auction feedback may induce learning and dependence in bids, which our static BNE model does not explicitly represent.

Equilibrium behavior is a maintained assumption of structural auction estimation, but verifying that SODA computes an approximate equilibrium for candidate primitives does not establish that human subjects play that equilibrium. The laboratory setting offers a valuable ground-truth evaluation: values are observed by the experimenter but withheld from estimation, so Table 3 measures recovery accuracy when an equilibrium-based estimator is applied to human bids. We interpret this as evidence of empirical recovery under a potentially imperfect behavioral approximation, not proof of equilibrium play. A comparison using later rounds only would directly assess sensitivity to learning. We will soften the claim of recovery “regardless of slot-quality decay,” given the higher errors under 11-10.

Reviewer_jj5MOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a Bayesian simulation method to estimate the value distribution of participants in a GSP auction. This method is compared with another one recently published, in three settings.

Critical strengths
  • I find the proposed idea simple and innovative
  • the authors managed to find some data to experiment with, which I believe is not that easy for this setting
Critical weaknesses

I enjoyed reading the paper, but I think it is not ready yet:

  • while I understand that in some settings it is not possible to provide general guaranties, the current version of the paper does not pass the bar. At least some argument of why the algorithm work, and what its failure mode are could be would be appreciated.
  • at least a special case where a property can be given would improve the paper contribution
  • Given the absence of theoretical contribution, the experimental section should be stronger: more synthetic experiments, error bars, more detailed algorithms and setup, reproducibility, compute times, ablations and variants, failure modes of naive approaches (see question below)
  • I believe the authors should take the time to explain the baseline they compare to in detail.
Questions for authors
  • Since we have access to the bid distribution of all participants, given their value, we can compute the best response of a participant by taking \argmax_bid \E_{competition} payoff(bid,competition_bids). So why do we need a simulation skim? I believe the paper does not explain that enough.
1 comment

Thank you for recognizing the idea as simple and innovative, and for noting the difficulty of obtaining auction data.

Why the algorithm works and when it fails (W1). The reasoning has four steps. (1) Under a strictly increasing equilibrium strategy β\beta, G(β(v))=F(v)G(\beta(v))=F(v) links bid and value distributions; monotonicity alone does not establish identification. (2) In the symmetric IPV GSP benchmark with known CTRs, linear scoring, and a strictly increasing symmetric BNE, the first-order condition yields v=ξ(b;G)v=\xi(b;G) (Kim & Pal, 2025). If F1F_1 and F2F_2 induce the same GG, then F1(ξ(b;G))=G(b)=F2(ξ(b;G))F_1(\xi(b;G))=G(b)=F_2(\xi(b;G)) on the relevant support, implying F1=F2F_1=F_2. For fixed Bernstein degree and support, an injective coefficient parameterization also identifies θ\theta. (3) If the true distribution belongs to this family, θ0\theta_0 uniquely minimizes Wp(Gθ,G0)W_p(G_\theta,G_0). (4) If computed equilibrium strategies deviate uniformly from their exact counterparts by at most δ\delta, coupling bids generated from identical value draws bounds the change in the Wasserstein criterion by δ\delta. A minimizer of the approximate criterion then achieves an objective value within 2δ2\delta of the exact minimum, not necessarily a parameter estimate within 2δ2\delta of the truth. Under identification, continuity, compactness, and vanishing sampling, simulation, solver, and optimization errors, standard minimum-distance arguments yield consistency (Bernton et al., 2019b). These are conditional results, not guarantees for the full BO-assisted SMC-ABC implementation.

Potential failure modes and diagnostic indicators include weak or absent identification (near-flat or multiple minima), solver inaccuracies (high best-response utility loss or unstable strategies), departures from equilibrium play (distorted value estimates), support misspecification (poor tail fit; Section 5.3), and surrogate screening that overlooks regions of the parameter space. These are indicators, not formal tests; low utility loss does not imply a uniform strategy-error bound. We will make these assumptions and limitations explicit.

Special case with guarantees (W2). Symmetric IPV GSP with known CTRs and linear scoring provides the requested special case. Within the specified monotone symmetric-equilibrium class, the inverse mapping establishes identification, and the coupling and minimum-distance arguments provide conditional error and consistency results. Gomes & Sweeney (2014) offer an analytical equilibrium characterization against which numerical strategies can be compared. More generally, best-response utility loss assesses approximate equilibrium quality in the discretized game, but not strategy distance to the exact BNE. We can report solver diagnostics.

Experiments (W3). We agree that error bars, ablations, runtime information, and failure cases would strengthen the evidence. Our three settings address different questions: synthetic recovery against a specialized baseline, structural recovery from human bids with known values, and a field-calibrated reserve-price counterfactual. These provide complementary proof-of-concept evidence, not exhaustive robustness validation. We can add hyperparameters and implementation details, report computational times and variability where available, and release code. We will discuss solver sensitivity and surrogate screening.

Baseline (W4). Kim & Pal (2025) adapt GPV-style two-step estimation to GSP with linear scoring: kernel estimates of bid CDFs and densities are used to invert the equilibrium first-order condition and recover pseudo-values, from which FF is estimated. Kernel boundary bias may contribute to the larger W2W_2 errors we observe (Hickman & Hubbard, 2015), but we cannot identify it as the cause without further analysis. We will expand the explanation of the baseline and our implementation choices.

Q1: Why not invert best responses to observed competing bids? This is a valid approach under equilibrium play and sufficient regularity: the optimality condition, given the competing-bid distribution, can sometimes be inverted to recover values. This is the principle behind GPV/Kim–Pal-style estimators. Our contribution is not that inversion is impossible in standard GSP, but that a mechanism-specific inverse mapping need not be derived or implemented. Instead, for each candidate value distribution, we compute equilibrium bids and compare their distribution to observed bids. A single best response against observed competition does not solve the equilibrium fixed point for a candidate value distribution. For reserve-price counterfactuals, bidders must likewise re-optimize against the new equilibrium bid distribution, not the old one. Our advantage is therefore a portable forward-simulation framework, subject to identification and equilibrium assumptions, rather than a claim that best-response inversion is invalid.

Reviewer_R8vYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks how a platform can recover advertisers' hidden values from bids in GSP auctions, where bids are strategic and not equal to values. Classical inversion methods work for first-price auctions but are mechanism-specific and hard to apply to GSP with quality scores.

The authors propose a simulation-based estimator. They represent candidate value distributions with flexible Bernstein polynomials, compute an approximate equilibrium for each candidate using an equilibrium-learning solver (SODA), simulate bids, and keep the candidate whose simulated bid distribution is closest to the observed one. A Bayesian-optimization-assisted sequential Monte Carlo procedure reduces the number of costly equilibrium computations.

They test the method on simulated benchmarks against an existing inversion-based estimator, on a GSP laboratory dataset with human subjects, and on a synthetic field-calibrated dataset for a reserve-price counterfactual. The method generally fits the bid distributions better than the benchmark and produces comparable revenue. The counterfactual suggests a 4% revenue gain over the prevailing reserve price.

The main weaknesses are that the method has no formal convergence or consistency guarantee, the equilibrium solver is used as a black box without sensitivity analysis, identification is assumed rather than demonstrated, the Bernstein distributions assume bounded support, and the experiments are small-scale. The paper is a promising practical pipeline, but its policy conclusions rest on assumptions it does not verify.

Critical strengths
  1. The paper reframes GSP value recovery as a forward-simulation problem, which makes structural estimation portable across mechanisms. The binding constraint in practice is not that GSP is hard to solve once, but that every mechanism change — quality scores, reserve prices, slot rules — requires a new analytical inversion. By treating equilibrium computation as a simulator and matching simulated to observed bid distributions, the paper replaces a mechanism-specific derivation with a mechanism-agnostic pipeline. This is consequential because the contribution is not a faster estimator but a reusable one: the same pipeline can be re-run when the auction rules change, which is precisely the constraint platforms face. Even though the current evidence is small-scale, the framing itself is a genuine contribution to the structural-estimation literature, and it is the main reason the paper is worth taking seriously.

  2. The validation is triangulated across three settings, including a rare ground-truth check with human subjects. The paper does not rely on a single simulation: it benchmarks against a specialized inversion-based estimator, tests on GSP laboratory data where the true values are observed but withheld from the estimator, and runs a field-calibrated counterfactual on reserve prices. The laboratory exercise is the most consequential part, because it asks whether the method can recover values from bids generated by real human subjects rather than from its own simulator. That gives the empirical case a breadth most methods papers lack, and it is the strongest reason the paper's practical claims are credible enough to support acceptance rather than rejection.

These two considerations — a portable framing and a tripartite validation — are what make the paper more than a black-box pipeline, and they are the factors that most strongly determine a positive recommendation.

Critical weaknesses
  1. Identification is assumed rather than established, and the experiments never test it in the setting that motivates the paper. The paper’s “identification conditions” (Results 1–4) amount to an injectivity assumption within a parametric Bernstein family: different candidate distributions must produce different bid distributions. That is a condition, not a demonstration that the data actually pin down the value distribution. The paper then lists censoring, endogenous participation, unobserved quality scores, and equilibrium selection as caveats — but these are not edge cases; they are defining features of real platform auctions, and each one breaks the mapping the estimator relies on. Worse, the experiments use linear scoring, a simple GSP lab setting, and a synthetic log-normal environment — exactly the cases where a specialized inversion or a simpler approach is available. The proprietary-quality-score case that motivates the paper is never tested. This is consequential because the paper’s central output — recovered values and the reserve-price counterfactual — is only as credible as the identification it assumes. If the conditions fail, the “recovered values” are simply the parameters that make simulated bids match, and the policy advice built on them has no structural interpretation. A revision cannot fix this with a paragraph of caveats; it would need an identification argument or a credible test.

  2. The computational pipeline is a black box with no error control, so the reader cannot tell whether the estimates are values or artifacts. The equilibrium is computed by SODA, but the paper never proves that SODA returns the symmetric Bayes–Nash equilibrium it defines, never reports solver error, and never shows how that error propagates into the recovered distribution. The ABC layer adds tolerance error, and the paper explicitly disclaims an unconditional convergence guarantee. This is consequential because the method’s entire claim is that simulation can substitute for analytical inversion; if the simulator’s accuracy is unknown, the substitution is not established. Change the solver’s initialization, tolerance, or equilibrium-selection rule, and the estimates may move — a possibility the paper does not test. Combined with the identification gap, this means the reader cannot distinguish a genuine recovery from a numerical coincidence, which is precisely what a structural-estimation paper must rule out.

These two weaknesses compound: identification is assumed, and the computational map from primitives to bids is unverified. That is why the paper’s policy conclusions are not yet established, and why the weaknesses are material rather than editorial.

Questions for authors

Three problems stand out in particular.

First, the equilibrium is computed through the black-box solver SODA, yet the paper never establishes that the solution SODA returns is in fact the equilibrium it defines. Nor does the paper offer any sensitivity analysis. Changing the initial values, the tolerance, or the equilibrium selection rule could all alter the results—and the paper never acknowledges this possibility.

Second, the paper asserts that its parameters are identified "under identification conditions," but it never demonstrates that these conditions hold in its data. If they do not hold, the estimates may bear no relation to the real world, and the counterfactual analyses and policy recommendations built on those estimates are rendered meaningless.

Third, the paper never proves that its estimator converges to the true value. Without that guarantee, a larger sample does not necessarily bring the estimate closer to the truth; the policy conclusions may be nothing more than a numerical coincidence.

The scalability issue is not fatal in itself. But taken together with the preceding three problems, it leaves the paper's claims of practical relevance without support.

The point is not whether any single flaw is fatal in isolation. The point is that the flaws compound: an unverified solver, unverified identification conditions, and unproven convergence together mean that the paper's core conclusions have not been established. This is not a defect that a disclaimer-style revision can repair—and it is the reason the paper cannot be accepted.

1 comment

Thank you for your detailed review. We appreciate your recognition of the forward-simulation approach as a reusable alternative to mechanism-specific inversion and of the three-part validation, particularly the laboratory ground-truth test. We address the main concerns below.

Identification (W1, Q2). We agree that Results 1–4 should distinguish an identification condition from evidence that it holds. For our symmetric IPV GSP benchmark with known CTRs, linear scoring, and a strictly increasing symmetric BNE, identification can be established through the equilibrium first-order condition. Monotonicity implies that bid and value ranks coincide. The first-order condition yields an inverse map v=ξ(b;G)v=\xi(b;G) in terms of the bid distribution GG (Kim & Pal, 2025). If two value distributions F1,F2F_1,F_2 generate the same GG under these assumptions, then F1(ξ(b;G))=G(b)=F2(ξ(b;G))F_1(\xi(b;G))=G(b)=F_2(\xi(b;G)) on the relevant support; hence F1=F2F_1=F_2. For fixed Bernstein degree and support, the linearly independent basis then implies distinct admissible coefficient vectors cannot represent the same CDF. We will make the assumptions and this benchmark argument explicit.

Our estimator does not establish identification by replacing inversion with simulation: it exploits identification where it holds without requiring the inverse map to be implemented. From a platform's perspective, its own scoring rule is known, even if proprietary to outsiders; nevertheless, identification for a more complex rule must be assessed separately. We agree that censoring, endogenous participation, unobserved bidder characteristics, and equilibrium selection can threaten identification. Nor do we claim that the present experiments validate recovery under proprietary nonlinear scoring. We will narrow the corresponding portability claims and identify such an experiment as future work.

Equilibrium solver (W2, Q1). Solver error can be related to our estimation criterion by coupling. Suppose, uniformly across candidate parameters θ\theta, the computed strategy β^θ\hat\beta_\theta differs from the exact strategy βθ\beta_\theta by at most δ\delta at every value. Using identical value draws under both strategies changes each simulated bid by at most δ\delta; hence their bid distributions differ by at most δ\delta in Wasserstein distance. By the triangle inequality, the resulting estimation criterion changes by at most δ\delta, uniformly in θ\theta. Thus an exact minimizer of the approximate criterion has an objective value at most 2δ2\delta above the minimum of the exact criterion. This is a bound on the objective value, not on the distance between estimated and true parameters; any optimization error must be added separately.

Importantly, the assumed uniform strategy-error bound is not implied by a small best-response utility loss. SODA is not an entirely unverified black box: its output can be evaluated by the utility gain from unilateral best responses, providing a diagnostic of approximate equilibrium quality in the discretized game (Bichler et al., 2023). We will clarify this distinction and report the available solver diagnostics. We restrict attention to the symmetric, strictly increasing equilibrium; other equilibrium-selection rules define different structural models. Sensitivity to initialization and solver tolerances remains an important limitation to assess.

Convergence (Q3). We do not claim unconditional consistency of the implemented BO-assisted SMC-ABC procedure. A conditional argument applies to the idealized minimum-Wasserstein estimator: identification, compactness of the parameter space, continuity of θ↦Gθ\theta\mapsto G_\theta in WpW_p, convergence of the empirical bid distribution to the population distribution, and uniformly vanishing solver and simulation errors imply consistency of approximate criterion minimizers, provided optimization error also vanishes (cf. Bernton et al., 2019b). ABC posterior concentration requires further conditions, including a vanishing tolerance (Bernton et al., 2019a), and surrogate screening must not invalidate the targeted ABC distribution. We will distinguish these conditional results from guarantees for our finite-computation implementation.

Bounded support and scale. Section 5.3 examines misspecification under an unbounded log-normal distribution: the upper tail is less accurately fitted, while estimated reserve prices and revenue change little in that experiment. This is evidence for that setting, not a general robustness guarantee. A transformation such as u=v/(1+v)u=v/(1+v) could accommodate unbounded values. We agree that the experiments are a proof of concept; candidate evaluations can be parallelized, and we can provide computation times to clarify practical scale.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.