acceptodds
Under review as a conference paper at ICLR 2027

Market-Price Learning under Endogenous Bid Censoring: Limits, Identification, and Correction

Abstract

In real-time bidding, a bidder's shading and pacing decisions depend on the distribution of the market price it must beat. The bidder learns this distribution from its own auction logs, which record every bid but reveal the price only when the bid wins. Statistically, this is estimation under dependent censoring in which the censoring variable, the bid, is always observed. Standard censored estimation relies on bid–price independence, but bids respond to the signals that drive prices, making the censoring endogenous. Without any assumption on this dependence, the logs identify the price margin only up to a sharp band whose width is the share of auctions lost with lower bids. This ambiguity yields an assumption-free minimax error floor. Under affiliation between the bid and the price, the band narrows to a closed-form interval, and standard censored estimation understates the price CDF. To correct this bias, we assume a concordance-ordered copula family. Given this family, the observed bids and win-region cells identify the dependence parameter and the price distribution through the largest bid. The sharp band then collapses to a point and the minimax floor vanishes over the same range. We fit neural bid and price distributions in two stages, first from the fully observed bids, then through a Gaussian-copula observed-data likelihood. For the finite model under standard regularity conditions, the two-stage estimator is consistent and asymptotically normal at the root- rate. On AuctionNet the correction cuts the largest gap of the censored network's CDF from 0.14 to 0.04 and closes about three quarters of its CRPS gap to the oracle. The gain persists when the bidding rule uses a market signal withheld from every estimator and when the coupling is planted outside the fitted family. It also reaches the decision layer, where the corrected bids cut the overshoot of a 10% target win rate on AuctionNet from 17.8 to 7.9pp.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
market price distribution, real-time bidding, endogenous censoring, dependent censoring, copula, identifiability, bid landscape forecasting

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2jPaOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to estimate market-price distributions from auction logs when prices are revealed only after a win. It first explains what cannot be recovered, then uses the always-observed bids and a copula dependence model to correct the bias. Experiments on three benchmarks assess prediction and bidding performance.

Critical strengths
  1. The paper makes good use of the fact that bids are always observed. It characterizes the remaining uncertainty and explains which added conditions allow recovery of the price distribution. (pp. 3–5, Theorem 1)
  2. The evaluation covers hidden signals, misspecified distribution families, and bid replay. On AuctionNet, the overshoot above the target win rate falls from 17.8 to 7.9 percentage points, showing a benefit for downstream decisions. (p. 8, Table 3)
Critical weaknesses
  1. Reliability remains unclear when most prices are hidden and bids closely track prices. The guarantee covers finite models under regularity conditions, not the actual network. The main distribution-error metric stops at the 90th percentile of bids, and error still rises to 0.075 under strong dependence. (pp. 5, 7, 21)
  2. Under misspecification, the diagnostics do not establish how much to trust the correction. All four candidate families are rejected in one iPinYou setting. The appendix also shows that a better likelihood can come with a larger price-distribution error. (pp. 8, 23)
Questions for authors
  1. How many low-bid wins are available to the actual network? If these samples are reduced, do the parameter estimates and corrected distributions remain stable?
  2. When all candidate families are rejected, what reliable conclusions can the logs still support? How much do the key bidding results vary across families?
3 comments
AuthorsOpenReview ↗

Response to the Review

We are grateful to the reviewer for valuable feedback and for highlighting the strengths of our work.

Q1. Low-bid wins and stability after reduction

Q1. How many low-bid wins are available to the actual network? If these samples are reduced, do the parameter estimates and corrected distributions remain stable?

(1) Low-bid win counts

There are 2,173 wins at the smallest bid of 7, and 8,725 wins in the bottom bid decile. AuctionNet has no wins at its smallest bid or in the bottom bid decile.

(2) Stability after reduction

On iPinYou at λ=0.5\lambda=0.5, reducing the smallest-bid rows by 98% left θ^\hat\theta nearly unchanged. The corrected model still had lower CDF error than the censored network. To address the reviewer's question, we reduced the iPinYou rows at the smallest bid and retrained with three seeds.

Fraction of smallest-bid rows retainedSmallest-bid wins in training split, seed meanChange in θ^\hat\theta from full data, per seedCorrected CDF errorCensored network
1.01,73100.0090.058
0.0239−0.016-0.016 to +0.003+0.0030.0120.052

On this stream, estimating θ\theta did not require retaining most of the rows at the smallest bid.

Q2, W2. Bidding across copula families and misspecification diagnostics

Q2. When all candidate families are rejected, what reliable conclusions can the logs still support? How much do the key bidding results vary across families?

W2 (summarized). Under misspecification, the diagnostics do not establish how much to trust the correction. All four families are rejected in one iPinYou setting, and a better likelihood can come with a larger price-distribution error.

(1) Bidding results across families

Gaussian and Frank give similar results on iPinYou. On AuctionNet, both improve over the censored network, with a larger gain for Gaussian.

pp, mean ± sd over 10 seedsCensored networkGaussianFrank
iPinYou λ=0.5\lambda=0.5, win-probability error6.1 ± 0.70.9 ± 0.31.0 ± 0.4
AuctionNet, overshoot of a 10% target win rate17.8 ± 1.67.9 ± 2.414.1 ± 1.4

The censored-network and Gaussian values are from Table 3.

(2) Conclusions after rejection

The logs still support the assumption-free sharp identified set (Proposition 1), which affiliation narrows to sharp bounds (Theorem 1(i)). In C2 of the general comment, we explain what remains supported when all families are rejected and why a likelihood gain does not establish price-margin accuracy.

W1. Strong dependence and CDF error above p90

W1 (summarized). Reliability remains unclear when most prices are hidden and bids closely track prices. The guarantee covers finite models under regularity conditions, not the actual network. The main error metric stops at the 90th percentile of bids, and error still rises to 0.075 under strong dependence.

(1) CDF error above p90

On the iPinYou rank mixture, correction reduced CDF error relative to the censored network between p90 and bmax⁡b_{\max}.

Mean CDF error, p90<k≤bmax⁡\mathrm{p90}<k\le b_{\max}λ=0.25\lambda=0.25λ=0.5\lambda=0.5λ=0.75\lambda=0.75
Censored network → correction0.073→0.0150.184→0.0670.306→0.249
Oracle0.0060.0060.006

We report errors up to bmax⁡b_{\max} because the log records no prices above it, and by Proposition 1 the CDF there is not identified without a tail assumption.

(2) Error under strong dependence

At λ=0.75\lambda=0.75, the correction still reduced CDF error up to p90 from 0.117 to 0.075, about 36% (Appendix B). Even given the true dependence, the copula-graphic recursion has a CDF error of 0.062 there (Figure 2, left). In C1 of the general comment, we address the reviewer's concern about the guarantee.

We are happy to answer further questions during the discussion period.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer 2jPa,

Thank you again for your review. The discussion period ends soon. Could you let us know whether our response addresses your concerns? It includes the row-reduction experiment you asked about in Q1 and the bidding results across copula families for Q2. We would be glad to answer any remaining questions.

The response addresses my concern, I decide to keep my positive score.

Reviewer_wQvKOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies market-price estimation from auction logs in which bids are always recorded, but prices are observed only when the bidder wins.The work analyzes identification and estimation bias under bid–price dependence and proposes a two-stage copula correction. Experiments show improvements in price-distribution estimation and bidding calibration, etc.

Critical strengths

The work makes useful use of information available in auction logs. Derived theory uses the always-observed bid and win-region cells to identify the dependence parameter and price margin within an assumed copula family. The sharp bounds also show what can be identified without this assumption, going beyond applying an existing dependent-censoring estimator.

The experiments help explain why the correction works. Independence controls, zero-dependence ablations, withheld signals, and non-Gaussian couplings help isolate the effect of dependence modeling. The fixed-bid experiments and family diagnostics help assess the method’s limitations.

Critical weaknesses

The theoretical guarantee does not directly cover the neural estimator used in the experiments. Theorem 2 assumes a fixed finite-dimensional model, while Proposition A.5 extends the result to finitely many feature values with a saturated fit. Neither result directly establishes consistency or root-n normality for the embedding networks used in the main experiments. The claim that the proposed model “carries the rate guarantee” needs further justification. Point identification also does not imply stable estimation when anchor cells have little probability mass or dependence is strong. The paper acknowledges this issue but does not quantify identification strength for the main conditional fits.

2.The diagnostics leave uncertainty about margin accuracy under misspecification. Appendix B.4 includes cases where likelihood improves while CDF error worsens. All candidate families also fail the observable-cell check on the iPinYou rank mixture at λ = 0.5. The correction still helps in the tested cases, but a practitioner without uncensored prices has little basis for judging its accuracy once all candidate families are rejected. It is noted that fixed-bid records offer validation below the exploration threshold, and on the other hand, recovery above it still relies on the assumed family.

Questions for authors
  1. Do the consistency or root-n guarantees cover the embedding-network estimator used in the main experiments? If so, what additional assumptions are needed? If not, please distinguish the finite-dimensional result from the neural implementation.
  2. On the main benchmarks, which feature groups have informative anchor cells, and how many observations fall in those cells? Could profile likelihoods or bootstrap estimates show how estimation variability changes as dependence increases? 3.How should practitioners interpret the corrected margin when all candidate copula families fail the observable-cell diagnostic? Is there evidence linking the diagnostic results or fixed-bid validation to margin accuracy?
2 comments
AuthorsOpenReview ↗

Response to the Review

We thank the reviewer for a thorough review and highly appreciate their interest in our work.

Q1, W1. Asymptotic guarantees for the neural network

Q1 (summarized). Do the consistency or root-n guarantees cover the embedding-network estimator in the main experiments? If so, under what additional assumptions? If not, please distinguish the finite-dimensional result from the neural implementation.

W1 (summarized). Theorem 2 assumes a fixed finite-dimensional model, and Proposition A.5 extends it to finitely many feature values with a saturated fit. Neither directly covers the embedding networks in the main experiments. The claim that the proposed model “carries the rate guarantee” needs further justification.

We have clarified the scope of the asymptotic guarantees and removed the claim that the neural estimator “carries the rate guarantee.” As explained in C1, extending these guarantees to the neural implementation requires additional representation and optimization conditions. Sections 1, 2.4, and 6 now clarify this distinction.

Q2, W1. Anchor-cell counts and variability across dependence strengths

Q2 (summarized). On the main benchmarks, which feature groups have informative anchor cells, and how many observations fall in them? Could profile likelihoods or bootstrap estimates show how estimation variability changes as dependence increases?

W1 (summarized). Point identification does not imply stable estimation when anchor cells have little mass or dependence is strong, and the paper does not quantify identification strength for the main conditional fits.

(1) Anchor cells and observation counts

At λ=0.5\lambda=0.5, iPinYou has 2,173 wins at the global minimum bid of 7. AuctionNet has none at its global minimum bid or in the bottom bid decile. Considering each of the 13 nonconstant iPinYou features separately, six have smallest-bid wins at every value.

Feature groupNumber of featuresValues without a smallest-bid win
Win at every value60%
Other features412–60%
High-cardinality features394–97%

The model estimates a shared θ\theta using observations across feature values, including values without smallest-bid wins. In the row-reduction experiment, removing 98% of the smallest-bid rows left θ^\hat\theta nearly unchanged (see R1 (Q1)).

(2) Variability across dependence strengths

On iPinYou, the bootstrap SDs of θ^\hat\theta and the corrected CDF did not increase monotonically with dependence.

Following the reviewer's suggestion, we assessed how estimation variability changes with dependence by retraining both stages, including θ\theta, on 50 bootstrap samples per setting that resample the training rows independently.

Settingθ^\hat\theta across bootstrap runs, mean (SD)Mean pointwise CDF SD
iPinYou, λ=0.25\lambda=0.250.395 (0.014)0.006
iPinYou, λ=0.5\lambda=0.50.740 (0.013)0.008
iPinYou, λ=0.75\lambda=0.750.883 (0.008)0.007
AuctionNet0.263 (0.030)0.013

Each fitted marginal CDF averages predictions over the same evaluation features. We computed its SD across runs at each price, then averaged these SDs over prices up to the 90th percentile of bids. Appendix B.3 (Tables 5 and 6) shows the anchor counts and bootstrap results.

Q3, W2. Interpreting the corrected margin when all families are rejected

Q3 (summarized). How should practitioners interpret the corrected margin when all candidate families fail the observable-cell diagnostic? Is there evidence linking the diagnostic or fixed-bid validation to margin accuracy?

W2 (summarized). Likelihood can improve while CDF error worsens (Appendix B.4), and all families fail the check on the iPinYou rank mixture at λ = 0.5. Without uncensored prices, a practitioner has little basis for judging accuracy once all families are rejected. Fixed-bid records validate below the exploration threshold, but recovery above it still relies on the assumed family.

(1) Interpreting rejection

Rejection can reflect a wrong family or shifted dependence and does not measure CDF error. The corrected margin should be interpreted alongside the fixed-bid check and the sharp identified set, which affiliation narrows (Theorem 1(i)). As the reviewer notes, recovery above tt relies on the assumed family. However, the fixed-bid experiment below provides evidence that discrepancy below tt can help assess accuracy above it.

(2) Evidence linking fixed-bid validation to margin accuracy

In the fixed-bid experiment of C2 (2), the maximum gap below tt between the corrected margin and the copula-free estimate from fixed-bid records had a Spearman correlation of 0.90 with CDF error above tt. Without true prices, practitioners can therefore check this gap in place of the unobservable error above tt to flag margins that are less accurate there.

We are happy to answer further questions during the discussion period.

Thanks for the response from the authors, which did make some points clear (e.g., some claims are removed). However, the contribution of the work did not increase much after the explanation. I am inclied to keep the same ratings.

Reviewer_krDBOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses market-price estimation from RTB logs, where the bidder's own bid is always recorded but the market price is observed only on wins, making censoring endogenous. It first derives the sharp set of price CDFs compatible with the observed logs under arbitrary bid–price dependence, shows that the width of this set is computable from the data, and pairs it with a matching minimax lower bound. Then, the authors prove that under Milgrom–Weber affiliation the conventional censored estimator understates the price CDF, and that within a copula family ordered by concordance, the fully observed bid distribution identifies both the copula parameter and the price CDF up to the highest logged bid. Moreover, they develop a two-stage estimator, i.e., an empirical bid margin plus a Gaussian-copula censored likelihood over neural hazards, with asymptotic guarantees. The authors validate their estimator on AuctionNet, iPinYou, and AuctionGym, showing that the correction closes most of the gap to an uncensored oracle and improves win-rate calibration and surplus estimation.

Critical strengths

S1. The identification theoretical analysis is thorough and complete end to end. Partial-identification bounds, the minimax floor, the affiliation-based bias direction, and the collapse of the band under a concordance-ordered copula are proved constructively as one coherent chain. The paper delineates precisely which obstacles are assumption-free impossibilities rather than adding another variant to the family of censored losses. S2. The experimental program interrogates the theory instead of just tabulating predictive metrics. Baselines that assume independence are shown to deteriorate as bid–price coupling is strengthened; the proposed estimator closes most of the distance to the uncensored oracle; benefits persist at the decision layer, in win-rate calibration and surplus; and the hidden-signal and out-of-family constructions are aimed squarely at the endogeneity mechanism the theory is about.

Critical weaknesses

W1. The data used for validation fall short of the realism the practical narrative invokes. All three testbeds are synthetic or semi-synthetic: AuctionNet is a calibrated simulator, AuctionGym is a re-implementation, and in the iPinYou data every retained row is a win at an almost constant logged bid, with bid variation restored by re-injecting a rank mixture across rows. Consequently, the correction has not yet been exercised on genuinely endogenous production logs, where limited support, heterogeneous dependence, and copula misspecification arise simultaneously instead of being planted one at a time. W2. Robustness to copula misspecification is demonstrated for a simpler estimator than the one recommended. The out-of-family experiments only concern the feature-free variant, and the planted setting even allows the selection procedure to choose the true family, while the deployed feature-conditional neural model is never stress-tested out of family. Moreover, on the iPinYou rank mixture the paper's own win-cell test rejects all four candidate families, yet the corrected CDF error still improves from 0.058 to 0.009, a gain that no out-of-family theory currently guarantees.

Questions for authors

Q1. Can the out-of-family findings be reproduced for the feature-conditional neural estimator that the paper actually deploys? And which portions of the asymptotic theory are meant to transfer to that estimator? Q2. On iPinYou, bids were reassigned through a rank mixture because the retained rows are all wins at nearly constant bids. Should the resulting numbers therefore be interpreted as robustness under controlled dependence, rather than as evidence about naturally logged bidder data?

2 comments
AuthorsOpenReview ↗

Response to the Review

We thank the reviewer for the positive assessment of our work and the valuable feedback

Q1, W2. Out-of-family results for the deployed neural estimator

Q1 (summarized). Can the out-of-family findings be reproduced for the feature-conditional neural estimator that the paper deploys? Which parts of the asymptotic theory transfer to it?

W2 (summarized). Out-of-family robustness is shown only for the feature-free variant, and the planted setting lets selection choose the true family, while the deployed neural model is never stress-tested out of family. On the iPinYou rank mixture, the win-cell test rejects all four families, yet CDF error still improves from 0.058 to 0.009, a gain no out-of-family theory currently guarantees.

(1) Out-of-family neural results

With the Gaussian family fixed, the deployed neural estimator reduced mean CDF error up to p90 on all three out-of-family streams.

Mean CDF error up to p90, mean ± sd over 10 seedsCensored networkProposedOracle
Rank mixture λ=0.5\lambda=0.50.058 ± 0.0060.009 ± 0.0030.004 ± 0.003
Planted Clayton0.075 ± 0.0050.021 ± 0.0050.004 ± 0.003
Planted Gumbel0.054 ± 0.0050.007 ± 0.0020.004 ± 0.003

The CDF improvement on these streams did not require a correctly specified copula family. We discuss win-cell rejection and CDF error in C2 of the general comment. In response to the reviewer's question, we added the planted-stream results to Appendix B.4 (Table 7).

(2) Scope of the asymptotic guarantees

Proposition A.5 gives consistency and root-nn normality under the representation and optimization conditions stated in C1, which the network trained in our experiments does not meet.

Q2, W1. Controlled dependence and production logs

Q2 (summarized). On iPinYou, bids were reassigned through a rank mixture because the retained rows are wins at nearly constant bids. Should the results be read as robustness under controlled dependence rather than as evidence about naturally logged bidder data?

W1 (summarized). All three testbeds are synthetic or semi-synthetic, and on iPinYou bid variation is restored by a rank mixture. The correction has not yet been exercised on endogenous production logs, where limited support, heterogeneous dependence, and copula misspecification arise together rather than being planted one at a time.

(1) Controlled dependence on iPinYou

We agree. The iPinYou results measure robustness under controlled dependence. For iPinYou, we kept the observed prices and reassigned the simulated bids from the DLF data through a rank mixture. Section 4 now clarifies how to interpret the iPinYou results.

(2) Dependence in AuctionNet

AuctionNet is a calibrated synthetic benchmark whose bidding agents use signals that also affect prices. The main experiment uses the simulator's original bid–price pairings. Its bid-replay gains in Table 3 were therefore obtained without planting the dependence.

We are happy to answer further questions during the discussion period.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer krDB,

We appreciate the time you spent on our paper. With the discussion period closing soon, could you tell us whether any of your concerns remain open? We ran the out-of-family experiment you requested on the deployed neural estimator (Appendix B.4, Table 7), and Section 4 now describes the iPinYou results as robustness under controlled dependence. Even a short note would help us.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

We are grateful to all reviewers for their constructive feedback and valuable suggestions. We conducted additional experiments to address the reviewers’ questions and uploaded a revised PDF with changes marked in blue. We first address two shared concerns, then list the manuscript changes with line numbers.

Reviewer labels. R1 = 2jPa, R2 = wQvK, R3 = krDB.

C1. Asymptotic guarantees and the neural implementation

Addresses R1 (W1), R2 (Q1, W1), and R3 (Q1).

Theorem 2 establishes consistency and root-nn asymptotic normality for the finite-dimensional estimator, and Proposition A.5 extends these guarantees to the saturated conditional model. Under the regularity conditions in Appendix A and with finitely many feature values, a network receives the same guarantees if it can represent every conditional table {G(b∣x)}\{G(b\mid x)\} and {F(k∣x)}\{F(k\mid x)\} and training reaches a maximizer at each stage. Its fit then coincides with the saturated fit. We have clarified the scope of these guarantees in Sections 1, 2.4, and 6.

C2. Interpreting the corrected margin when all families are rejected

Addresses R1 (Q2, W2), R2 (Q3, W2), and R3 (W2).

(1) Does rejection of every family mean the correction failed? R1 (W2), R2 (W2), R3 (W2)

Not necessarily. Rejection by the win-cell check signals a wrong family or shifted dependence but does not measure CDF error (Section 4). On the iPinYou rank mixture at λ=0.5\lambda=0.5, the win-cell check rejected all four families for the estimator without features. On the same stream, the proposed model reduced CDF error from the censored network's 0.058 to 0.009, near the oracle's 0.004.

(2) Can fixed-bid records assess accuracy without true prices? R2 (Q3, W2)

Fixed-bid records validate the margin up to tt. Their discrepancy from the corrected margin was associated with CDF error above tt in an additional experiment. We agree that under copula misspecification, likelihood alone does not establish margin accuracy.

Suppose some auctions are selected independently of prices and bidding inputs, and a fixed bid tt is used in each. For each k≤tk\le t, the fraction of these auctions won at a recorded price at most kk estimates F(k)F(k) without a copula (Proposition A.3). The maximum absolute gap between this estimate and the corrected margin over prices up to min⁡(t,p90)\min(t,\mathrm{p90}) can be computed without true prices. Across 64 settings on eight streams with known prices, this gap had a Spearman correlation of 0.90 with CDF error above tt. Practitioners can therefore check this gap in place of the unobservable error above tt to flag margins that are less accurate there.

(3) What do the logs still support? R1 (Q2)

The assumption-free sharp identified set computed from observed cells (Proposition 1). Under affiliation, Theorem 1(i) narrows this set to sharp bounds without choosing a copula family. Reporting the corrected margin alongside these bounds distinguishes the range supported by the log from the point estimate that relies on the assumed family.

Manuscript changes

The revised PDF makes the following changes in response to the reviews.

LocationChangeRelated questions
Section 1 (lines 103–104), Section 2.4 (line 269), Section 6 (lines 491–492)Distinguishes the finite-dimensional maximizer and saturated fit from the neural implementation; replaces “carries the rate guarantee”R1 (W1), R2 (Q1, W1)
Section 4 (lines 311–313)Describes the iPinYou results as robustness under controlled dependenceR3 (Q2)
Appendix B.4 (lines 1254–1256), Table 7Adds out-of-family experiments for the deployed neural estimatorR3 (Q1, W2)
Appendix B.3 (lines 1178–1183), Tables 5 and 6Adds iPinYou smallest-bid win counts and bootstrap variabilityR2 (Q2, W1)

Separately, to aid understanding, we added a one-line argument for F^{\\#}\le F_L in Section 2 to connect Proposition 2 with Theorem 1(i). We also clarified the affiliation discussion in the abstract and Sections 1 and 6.

AuthorsOpenReview ↗

Summary of the Discussion

Dear Area Chair and Reviewers,

Thank you for your time. We summarize the discussion below.

Contribution.

For auction logs where the price is seen only on wins, we give the sharp identified set (Proposition 1), sharp bounds under affiliation (Theorem 1(i)), and point identification within a copula family from the always-observed bid (Theorem 1(ii)). A two-stage estimator corrects the bias and is root-nn consistent in the finite-dimensional model (Theorem 2).

Strengths noted by the reviewers.

All three reviewers rated the paper highly for both presentation and informativeness.

  • R3: "delineates precisely which obstacles are assumption-free impossibilities rather than adding another variant to the family of censored losses."
  • R2: "going beyond applying an existing dependent-censoring estimator."
  • R1: "characterizes the remaining uncertainty and explains which added conditions allow recovery of the price distribution."

Additional evidence added during rebuttal.

  1. Out-of-family test of the deployed neural estimator (R3). Mean CDF error up to p90 fell on all three streams, for example from 0.058 to 0.009 (Table 7).
  2. Accuracy when every family is rejected (R1, R2, R3). The fixed-bid gap below tt had a Spearman correlation of 0.90 with CDF error above tt.
  3. Reliance on low-bid wins (R1, R2). Removing 98% of the smallest-bid rows left θ^\hat\theta nearly unchanged. Appendix B.3 adds smallest-bid win counts and bootstrap variability (Tables 5 and 6).

Manuscript changes are marked in blue in the revised PDF.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.