Pith. sign in

REVIEW 2 major objections 5 minor 51 references

The paper proves that optimal bidding across heterogeneous ad auctions reduces to equalizing marginal cost across channels, and shows a 3% GMV lift in online tests.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

HOB equalizes marginal cost across heterogeneous auction channels and uses a zero-inflated exponential win-price model for first-price auctions with organic traffic, reporting a 3.0% GMV lift in online A/B tests.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Correct but mostly an identity; the deployed system is real, but the global-optimality claim rests on an unvalidated power-law fit. the 2 major comments →

arxiv 2510.15238 v2 pith:TOUFLVK2 submitted 2025-10-17 cs.GT cs.IRcs.LG

HOB: A Holistically Optimized Bidding Strategy under Heterogeneous Bidding Environments

Qi Li , Wendong Huang , Qichen Ye , Wutong Xu , Cheems Wang , Wei Yuan , Miao Xu , Zhiyu Mou
show 4 more authors
Guan Wang Rongquan Bai Chuan Yu Jian Xu
This is my paper
classification cs.GT cs.IRcs.LG MSC 91B26
keywords auto-biddingfirst-price auctionbid shadingmarginal cost equalizationwinning price distributionzero-inflated exponentialorganic trafficheterogeneous auction channels
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Modern advertisers run one campaign across channels with different auction rules — second-price auctions, first-price auctions with uniform bids, first-price auctions with per-impression shading, plus organic traffic that can be won for free. This paper argues that the whole multi-channel problem collapses to a single control rule: pick bids so that the marginal cost of acquiring value is the same in every channel. To make that rule practical, it derives the marginal cost for each channel type, including the result that in a first-price channel with shading the marginal cost equals the global control parameter exactly. Local optimality is handled by modeling the winning price with a zero-inflated exponential distribution, which captures the large mass of impressions winnable with a zero bid. The delivered system, HOB, reports a 3.0% lift in GMV in large-scale online A/B tests on a commercial DSP.

Core claim

The central discovery is a pair of marginal-cost identities that make heterogeneous channels commensurable. In a first-price auction where the bid on each impression is chosen to maximize (eta*v_i - bid) * p_i(bid), the aggregate marginal cost C'(eta)/V'(eta) equals eta exactly (Theorem 3.2), regardless of the mixture of free and paid wins; the first-order condition of the surplus maximization plus the envelope theorem forces the ratio. In a uniform first-price channel the marginal cost is eta + V(eta)/V'(eta), and in a second-price channel it is simply eta. The paper then applies the classical principle that an interior optimum equalizes marginal cost across all channels and derives a close

What carries the argument

The load-bearing objects are the ZIE winning-price distribution and the marginal-cost identities. ZIE, a zero-inflated exponential with parameters (pi, lambda), places a point mass at zero win price (organic traffic) and an exponential tail, making the surplus function (V - x)F(x) strictly unimodal so the optimal bid is found by a few iterations of golden-section search. The MC identities convert each channel's bidding parameter into a common currency: MC=eta for SPA, MC=eta + V(eta)/V'(eta) for uniform FPA, and MC=eta for shaded FPA. The MCA module sets the uniform-FPA parameter via the power-law correction eta3 = eta/(1+1/b), equalizing all channels at a shared marginal-cost target.

Load-bearing premise

The alignment for uniform first-price channels assumes the total value function follows a power law V(eta3) = a * eta3^b in the operating region; if that shape is wrong, the correction eta3 = eta/(1+1/b) misaligns the channel and the claimed global optimality is not realized.

What would settle it

Regress realized value on the uniform-FPA bid multiplier across a range of eta3 values in log-log space using traffic replay; if the curve is not linear (the slope b changes with eta3), the closed-form correction is misspecified. Alternatively, measure realized marginal cost C'(eta)/V'(eta) on the FPA+u channel in production: if it departs from eta while the SPA and FPA+nu channels sit at eta, the power-law calibration is off.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Any interior optimum of the multi-channel constrained bidding problem must equalize marginal cost across channels; HOB's per-channel MC formulas supply the operational values needed to enforce that condition.
  • In a first-price channel with per-impression shading, the global control parameter eta is exactly the marginal cost of value, so the same eta that satisfies campaign constraints also sets the correct cross-channel tradeoff.
  • Modeling winning price as ZIE yields a strict unimodal surplus, guaranteeing a unique, cheaply computed optimal bid per impression and a closed-form decision rule for when to bid zero.
  • Equalizing MC rather than applying the same bidding multiplier everywhere recovers clicks and GMV: on the YOYI replay HOB improves clicks by 6.18% over the unified-multiplier baseline at the same cost.
  • In live traffic on a commercial DSP, HOB raises GMV by 3.0% and cost by 3.1% while keeping ROAS compliance, and removing the MCA module degrades the outcome, confirming the alignment step carries the gain.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This extends beyond the paper: Theorem 3.2's MC=eta identity is derived from the first-order condition of surplus maximization, which does not actually depend on ZIE's specific shape; the same equality should hold for any smooth, single-peaked winning-price model, so the alignment rule may be portable to other distributional priors.
  • This extends beyond the paper: the power-law assumption V = a*eta^b is testable in the fitted traffic; plotting log V against log eta should be linear. If it is not, the fixed exponent 1/b could be replaced by a locally estimated elasticity, giving a nonparametric version of the same alignment.
  • This extends beyond the paper: the value-weighted perspective suggests the same MCA module can absorb new channel types (e.g., auctions with reserve prices or multi-slot outcomes) by deriving their MC identity and slotting it into the shared-target loop.
  • This extends beyond the paper: since the online gain concentrates in cost saving on organic traffic, advertisers with high organic share and heterogeneous channel mix likely benefit most; a targeted A/B split by organic-share stratum would test that prediction.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes HOB, a bidding framework for advertisers operating across heterogeneous auction channels: SPA with uniform bids, FPA with uniform bids (FPA+u), and FPA with non-uniform bid shading (FPA+nu), in a setting where organic traffic can be won with zero bid. The winning price distribution is modeled as a zero-inflated exponential (ZIE). For FPA+nu, per-impression surplus maximization is solved, and Theorem 3.1 proves unimodality. The paper then derives marginal cost formulas: MC=eta_s for SPA, MC=eta_f+V/V' for FPA+u, and MC=eta for FPA+nu (Theorem 3.2). A marginal cost alignment (MCA) module equalizes MC across channels by adjusting the per-channel control parameters. For FPA+u, the V/V' term is approximated by fitting a power law V(eta3)=a*eta3^b, leading to Eq. (16), eta3=eta/(1+1/b). Experiments include synthetic data, YOYI offline logs, industrial log replay, and online A/B testing, reporting a 3.0% GMV lift on a commercial DSP. The global-optimality claim rests on Lemma 1, imported from prior work, and on the correctness of the power-law approximation.

Significance. If the MCA module is correct, the paper offers a practical, scalable operationalization of marginal-cost equalization in a real industrial multi-channel autobidding system. The ZIE model for FPA with organic traffic is a sensible and well-motivated modeling choice, and Theorem 3.1's proof is sound. Theorem 3.2's MC=eta identity, while a direct consequence of the first-order condition, is a useful characterization. The paper also provides rare evidence from a deployed system with online A/B results, which strengthens the practical contribution. However, the power-law approximation underlying the FPA+u alignment rule is unvalidated and structurally questionable when organic traffic yields V(0)>0; this is load-bearing for the global-optimality claim. The regularity conditions for Lemma 1 are also not stated.

major comments (2)
  1. [Section 3.3, Eq. (16)] The FPA+u alignment rule rests on the power-law assumption V(eta3)=a*eta3^b, with no goodness-of-fit or residual diagnostics reported. In the organic-paid setting, the ZIE win probability implies V(0)=sum_i v_i pi_i >0, so a pure power law is structurally inconsistent with the motivating setting; V/V' need not equal eta3/b, and the rule eta3=eta/(1+1/b) does not secure MC equalization. The paper itself notes that the exact V/V' is computable by traffic replay; if the offline Table 2 uses exact replay rather than Eq. (16), it does not validate the deployed approximation. Please report diagnostics for the power-law fit, or use the exact V/V' online, and re-evaluate the global-optimality claim.
  2. [Section 3.3, Lemma 1] The global-optimality claim relies on Lemma 1, but the proof is deferred to [1] and refers to 'Assumption 3.1' without reproducing it. The paper should state those regularity conditions and verify that they hold for the ZIE-based FPA+u and FPA+nu channels (e.g., concavity/smoothness of V and C, interiority). Without this, the statement that HOB 'provably secures bidding efficiency' and 'achieves globally optimal allocation' is not self-contained.
minor comments (5)
  1. [Section 3.3, Theorem 3.2] The MC=eta result is a direct algebraic consequence of the first-order condition in Eq. (12) and the definition of the aggregate value/cost functions; it should be presented as an operational characterization rather than an independent prediction. This does not affect correctness, but the wording 'rigorous derivation' overstates the novelty.
  2. [Section 3.3, Eq. (8)] The SPA marginal cost formula MC=eta_s is asserted without derivation. Please provide a derivation or a precise citation to the corresponding result in [1], since it is used in the MCA alignment.
  3. [Section 4.4, Table 4] The online A/B results report relative lifts but no confidence intervals, p-values, or test details beyond bucket assignment. Please add statistical significance information, especially for the key GMV and Cost metrics.
  4. [Section 4.2, organic-traffic experiment] The organic traffic is simulated by adding zero-mean Gaussian noise with std=70% to YOYI winning prices and clipping at zero. This synthetic transformation may not reproduce the real organic-win structure; please discuss its limitations and how the conclusions transfer to the real organic setting.
  5. [Appendix A.1, Algorithm 2] Algorithm 2 contains a stray top-level call to OptimalBidCalculation after the function definitions (line 19), which appears to be a formatting artifact. Also, Algorithm 1 sets N_iter=10 while the text states 'typically 6 iterations'; please reconcile.

Circularity Check

0 steps flagged

No significant circularity: Theorem 3.2 is an envelope-theorem identity with a self-contained proof, Lemma 1 is an external KKT result, and the power-law MCA rule is an unvalidated approximation, not a circular prediction.

full rationale

The central derivation chain is not circular. Lemma 1 is imported from an external reference [1] (Aggarwal et al. 2024) with no author overlap; it is a standard KKT marginal-cost-equalization condition and is not fitted in this paper. Theorem 3.2 is an envelope-theorem identity: given bid* = argmax (v_i*eta - bid_i)*p_i(bid_i), the first-order condition forces C'(eta) = eta*V'(eta), so MC(eta)=C'/V'=eta. This is a mathematical consequence of the stated optimization problem, not a prediction fitted to data; even though eta is the Lagrange-multiplier-derived control parameter, the theorem's proof does not assume its own conclusion and derives the equality from the FOC and chain rule. The MCA rule for FPA+u, Eq (16), is derived algebraically from the assumed power-law V(eta3)=a*eta3^b whose parameters are fitted from FPA data; this is an unvalidated functional-form approximation (especially questionable with organic traffic giving V(0)>0), but it is not circular: the fitted b is used in a control law, not renamed as an independent prediction or as evidence for the theory. The main weakness is that the synthetic-data experiment in Section 4.1 generates winning prices from the same ZIE process the paper assumes, making Table 1 a self-consistency check rather than independent validation; this weakens the empirical support for the ZIE choice but does not make the paper's derivation circular. Overall, the theoretical claims are either externally sourced or proven from explicit assumptions, and the practical approximation is a correctness risk, not a circular step.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central derivation relies on: (1) MC equalization from KKT theory (cited to [1]), (2) the ZIE distributional prior, and (3) the power-law value curve for the uniform-FPA channel. The latter is a fitted approximation used to make the MCA module a single coefficient multiplication; the fit parameters a,b are advertiser-specific and not reported. No new entities are introduced.

free parameters (3)
  • Power-law scale a = not reported
    In the value function V(eta3)=a*eta3^b fitted per advertiser from FPA data; used in Eq (16) to derive eta3=eta/(1+1/b).
  • Power-law exponent b = not reported
    Fitted per advertiser from FPA data; determines the MCA adjustment eta3 = eta/(1+1/b). This is an advertiser-specific fitted parameter, a core degree of freedom in the MCA module.
  • ZIE parameters (pi_i, lambda_i) per impression = learned by MLE
    Parameters of the winning price distribution used to compute surplus and optimal bid; trained on production/YOYI logs. Not adversarially tuned but are fitted model parameters.
axioms (5)
  • domain assumption Interior optimal solution and MC equalization (Lemma 1)
    Section 3.3: 'If an optimal bidding strategy exists and is an interior solution... then this strategy is unique and satisfies the marginal cost equalization condition.' Proof cited to [1]; assumes regularity conditions from [1].
  • ad hoc to paper Winning price follows ZIE distribution
    Section 3.2.1: models winning price with zero-inflation probability pi and exponential rate lambda; used for F in Theorem 3.1 and Theorem 3.2.
  • ad hoc to paper Value function is a power law V(eta3)=a*eta3^b in a neighborhood of the current eta3
    Section 3.3, Eq (16) derivation: 'Considering that the value function follows a power-law form...' This is the basis of Eq (16); no justification beyond empirical fit.
  • domain assumption Budget exhaustion effects are negligible (cost per request minuscule relative to budget)
    Section 3.3: 'Since the cost per request in real-world industrial scenarios is minuscule relative to the overall budget, we temporarily neglect the effects of budget exhaustion.' This formalizes the problem with only an ROI constraint.
  • domain assumption Independence of impressions / regularity of distribution model
    Implicit in the surplus maximization factorization across impressions; also the Lagrangian relaxation relies on independent per-impression choices.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of HOB: A Holistically Optimized Bidding Strategy under Heterogeneous Bidding Environments." pith.science (2026). https://pith.science/paper/TOUFLVK2

@misc{pith2026251015238,
  author       = {Pith},
  title        = {Pith review of: HOB: A Holistically Optimized Bidding Strategy under Heterogeneous Bidding Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TOUFLVK2}},
  note         = {Machine review of arXiv:2510.15238}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Optimizing a single advertising campaign across heterogeneous channels is a central challenge in industrial autobidding. Auction mechanisms vary across channels in ranking rules (pure eCPM vs. UE-augmented scoring), pricing formats (first- vs. second-price), and bidding conventions (uniform vs. non-uniform), while advertisers impose shared campaign-level constraints. We propose HOB, which makes marginal cost (MC) computable and alignable across heterogeneous channels, especially for first-price auctions (FPA) with organic-paid coexistence, where existing bidding formulations do not yield a practical aligned MC form. At the global level, HOB derives channel-specific MC forms and coordinates disparate channels through a shared MC target. At the local level, HOB models free-win probability and winning-price uncertainty with a zero-inflated exponential distribution, yielding an efficient surplus-optimal bidding strategy for non-uniform first-price auctions. We show that any interior optimum satisfies MC equalization across channels. Experiments on a controlled offline benchmark, industrial log replay, and large-scale online A/B tests demonstrate that HOB consistently delivers significant performance gains. Deployed on a large-scale commercial DSP, HOB delivers a 3.0% lift in GMV while maintaining return on advertising spend (ROAS) constraints.

Figures

Figures reproduced from arXiv: 2510.15238 by Cheems Wang, Chuan Yu, Guan Wang, Jian Xu, Miao Xu, Qichen Ye, Qi Li, Rongquan Bai, Wei Yuan, Wendong Huang, Wutong Xu, Zhiyu Mou.

Figure 1
Figure 1. Figure 1: Illustration of the procedure. Our work focuses on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between the actual CDF (black) and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The architecture of our proposed HOB. While the Optimal Bid Calculation module (left) computes a locally optimal [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: HOB’s improvements on different budget levels. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Analysis on different channel proportions. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 9 linked inside Pith

  1. [1]

    Gagan Aggarwal, Andres Perlroth, Ariel Schvartzman, and Mingfei Zhao. 2024. Platform Competition in the Autobidding World. arXiv:2405.02699 [cs.GT] https://arxiv.org/abs/2405.02699

    Pith/arXiv arXiv 2024

  2. [2]

    Mohammad Akbarpour and Shengwu Li. 2020. Credible Auctions: A Trilemma. Econometrica88 (2020)

    2020

  3. [3]

    Dirk Bergemann, Benjamin Brooks, and Stephen Morris. 2017. First-price auc- tions with general information structures: Implications for bidding and revenue. Econometrica85, 1 (2017), 107–143

    2017

  4. [4]

    Jason Bigler. 2019. Rolling out first price auctions to Google Ad Manager part- ners. https://blog.google/products/admanager/rolling-out-first-price-auctions- google-ad-manager-partners/

    2019

  5. [5]

    Nikolay Borissov, Dirk Neumann, and Christof Weinhardt. 2010. Automated bidding in computational markets: an application in market-based allocation of computing services.Autonomous Agents and Multi-Agent Systems21, 2 (2010), 115–142

    2010

  6. [6]

    H. Cai, K. Ren, W. Zhag, K. Malialis, and D. Guo. 2017. Real-Time Bidding by Reinforcement Learning in Display Advertising.ACM(2017)

    2017

  7. [7]

    Bhaskar Chakravorti, William W Sharkey, Yossef Spiegel, and Simon Wilkie

  8. [8]

    V Chari and Robert Weber. 1992. How the U.S. Treasury Should Auction Its Debt. https://kylewoodward.com/blog-data/pdfs/references/chari+weber- quarterly-review-1992A.pdf

    1992

  9. [9]

    Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. 2021. Decision transformer: Reinforcement learning via sequence modeling.Advances in neural information processing systems34 (2021), 15084–15097

    2021

  10. [10]

    Peter Cramton. 2004. Competitive bidding behavior in uniform-price auction markets. In37th Annual Hawaii International Conference on System Sciences, 2004. Proceedings of the. IEEE, 11–pp

    2004

  11. [11]

    Easley David and Kleinberg Jon. 2010. Networks, Crowds, and Markets: Rea- soning About a Highly Connected World. https://www.cs.cornell.edu/home/ kleinber/networks-book/networks-book-ch09.pdf

    2010

  12. [12]

    Yuan Deng, Negin Golrezaei, Patrick Jaillet, Jason Cheuk Nam Liang, and Vahab Mirrokni. 2023. Multi-channel autobidding with budget and roi constraints. In International Conference on Machine Learning. PMLR, 7617–7644

    2023

  13. [13]

    Mirrokni, Hanrui Zhang, and Song Zuo

    Yuan Deng, Jieming Mao, V. Mirrokni, Hanrui Zhang, and Song Zuo. 2022. Effi- ciency of the First-Price Auction in the Autobidding World.ArXivabs/2208.10650 (2022)

    Pith/arXiv arXiv 2022

  14. [14]

    Yuan Deng, Jieming Mao, Vahab Mirrokni, and Song Zuo. 2021. Towards Efficient Auctions in an Auto-bidding World. (2021)

    2021

  15. [15]

    Paulo Fagandini and Ingemar Dierickx. 2023. Computing profit-maximizing bid shading factors in first-price sealed-bid auctions.Computational Economics61, 3 (2023), 1009–1035

    2023

  16. [16]

    Tong Geng, Fangzhou Sun, Di Wu, Wei Zhou, Harikesh Nair, and Zhangang Lin

  17. [17]

    Getintent. 2017. RTB Auctions: Fair Play? https://blog.getintent.com/rtbauctions- fair-play-3b372d505089

    2017

  18. [18]

    Djordje Gligorijevic, Tian Zhou, Bharatbhushan Shetty, Brendan Kitts, Shengjun Pan, Junwei Pan, and Aaron Flores. 2020. Bid shading in the brave new world of first-price auctions. InProceedings of the 29th ACM International Conference on Information & Knowledge Management. 2453–2460

    2020

  19. [19]

    Zhen Gong, Lvyin Niu, Yang Zhao, Miao Xu, Haoqi Zhang, Zhenzhe Zheng, Zhilin Zhang, Rongquan Bai, Chuan Yu, Jian Xu, et al. 2023. MEBS: Multi-task End-to-end Bid Shading for Multi-slot Display Advertising. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 4588–4594

    2023

  20. [20]

    Jiayan Guo, Yusen Huo, Zhilin Zhang, Tianyu Wang, Chuan Yu, Jian Xu, Bo Zheng, and Yan Zhang. 2024. Generative auto-bidding via conditional diffusion modeling. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 5038–5049

    2024

  21. [21]

    Xiaotian Hao, Zhaoqing Peng, Yi Ma, Guan Wang, Junqi Jin, Jianye Hao, Shan Chen, Rongquan Bai, Mingzhou Xie, Miao Xu, et al. 2020. Dynamic knapsack op- timization towards efficient multi-channel sequential advertising. InInternational Conference on Machine Learning. PMLR, 4060–4070

    2020

  22. [22]

    Yue He, Xiujun Chen, Di Wu, Junwei Pan, Qing Tan, Chuan Yu, Jian Xu, and Xiaoqiang Zhu. 2021. A unified solution to constrained bidding in online display advertising. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2993–3001

    2021

  23. [23]

    Ali Hortaçsu, Jakub Kastl, and Allen Zhang. 2018. Bid shading and bidder surplus in the us treasury auction system.American Economic Review108, 1 (2018), 147–169

    2018

  24. [24]

    Hovaness

    B. Hovaness. 2018. Sold for more than you should have paid. https://www. heartsscience.com/sold-for-more-than-you-should-have-paid/

    2018

  25. [25]

    Jiani Huang, Zhenzhe Zheng, Yanrong Kang, and Zixiao Wang. 2024. From Sec- ond to First: Mixed Censored Multi-Task Learning for Winning Price Prediction. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 295–303

    2024

  26. [26]

    Pingjun Jiang. 2025. Automated bidding vs manual bidding strategies in search engine marketing: a keyword efficiency perspective.Journal of Marketing Ana- lytics13, 1 (2025), 82–95

    2025

  27. [27]

    Junqi Jin, Chengru Song, Han Li, Kun Gai, Jun Wang, and Weinan Zhang. 2018. Real-Time Bidding with Multi-Agent Reinforcement Learning in Display Adver- tising.ACM(2018)

    2018

  28. [28]

    Niklas Karlsson and Qian Sang. 2021. Adaptive Bid Shading Optimization of First-Price Ad Inventory. In2021 American Control Conference (ACC). 4983–4990. doi:10.23919/ACC50511.2021.9482665

    arXiv 2021

  29. [29]

    Bernhard Kasberger and Karl H Schlag. 2024. Robust bidding in first-price auctions: How to bid without knowing what others are doing.Management Science70, 7 (2024), 4219–4235

    2024

  30. [30]

    Randall Lewis and Jeffrey Wong. 2022. Incrementality Bidding and Attribution. arXiv:2208.12809 [cs.LG] https://arxiv.org/abs/2208.12809

    Pith/arXiv arXiv 2022

  31. [31]

    Ningyuan Li, Zhilin Zhang, Tianyan Long, Yuyao Liu, Rongquan Bai, Yurong Chen, Xiaotie Deng, Pengjie Wang, Chuan Yu, Jian Xu, et al. 2025. Beyond Adver- tising: Mechanism Design for Platform-Wide Marketing Service’QuanZhanTui’. A vailable at SSRN 5190237(2025)

    2025

  32. [32]

    R Preston McAfee and John McMillan. 1987. Auctions and bidding.Journal of economic literature25, 2 (1987), 699–738

    1987

  33. [33]

    Daisuke Moriwaki, Yuta Hayakawa, Isshu Munemasa, Yuta Saito, and Akira Matsui. 2020. Unbiased Lift-based Bidding System. arXiv:2007.04002 [cs.LG] https://arxiv.org/abs/2007.04002

    Pith/arXiv arXiv 2020

  34. [34]

    Zhiyu Mou, Yusen Huo, Rongquan Bai, Mingzhou Xie, Chuan Yu, Jian Xu, and Bo Zheng. 2022. Sustainable Online Reinforcement Learning for Auto-bidding. arXiv:2210.07006 [cs.LG]

    Pith/arXiv arXiv 2022

  35. [35]

    Renato Paes Leme, Balasubramanian Sivan, and Yifeng Teng. 2020. Why do competitive markets converge to first-price auctions?. InProceedings of The Web Conference 2020. 596–605

    2020

  36. [36]

    Shengjun Pan, Brendan Kitts, Tian Zhou, Hao He, Bharatbhushan Shetty, Aaron Flores, Djordje Gligorijevic, Junwei Pan, Tingyu Mao, San Gultekin, et al. 2020. Bid shading by win-rate estimation and surplus maximization.arXiv preprint arXiv:2009.09259(2020)

    Pith/arXiv arXiv 2020

  37. [37]

    Yanlin Qu, Ravi Kant, Yan Chen, Brendan Kitts, San Gultekin, Aaron Flores, and Jose Blanchet. 2024. Double distributionally robust bid shading for first price auctions.arXiv preprint arXiv:2410.14864(2024)

    Pith/arXiv arXiv 2024

  38. [38]

    Kan Ren, Jiarui Qin, Lei Zheng, Zhengyu Yang, Weinan Zhang, and Yong Yu. 2019. Deep landscape forecasting for real-time bidding advertising. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 363–372

    2019

  39. [39]

    Marco Slikker. 2007. Bidding for surplus in network allocation problems.Journal of Economic Theory137, 1 (2007), 493–511

    2007

  40. [40]

    Sarah Sluis. 2017. Big Changes Coming To Auctions, As Exchanges Roll The Dice On First-Price. https://www.adexchanger.com/platforms/big-changes-coming- auctions-exchanges-roll-dice-first-price/

    2017

  41. [41]

    Fransisca Susan, Negin Golrezaei, and Okke Schrijvers. 2023. Multi-platform budget management in ad markets with non-ic auctions.arXiv preprint arXiv:2306.07352(2023)

    Pith/arXiv arXiv 2023

  42. [42]

    Nair, and Carlos Carrion

    Caio Waisman, Harikesh S. Nair, and Carlos Carrion. 2025. Online Causal Inference for Advertising in Real-Time Bidding Auctions.Marketing Science44, 1 (2025)

    2025

  43. [43]

    Jian Xu, Xuhui Shao, Jianjie Ma, Kuang-chih Lee, Hang Qi, and Quan Lu. 2016. Lift-based bidding in ad selection. InProceedings of the Thirtieth AAAI Conference on Artificial Intelligence(Phoenix, Arizona)(AAAI’16). AAAI Press, 651–657

    2016

  44. [44]

    Wei Zhang, Shuai Yuan, and Jun Wang. 2016. User Response Learning for Directly Optimizing Campaign Performance in Display Advertising.arXiv preprint arXiv:1404.1898(2016)

    Pith/arXiv arXiv 2016

  45. [45]

    Dongwei Zhao, Audun Botterud, and Marija Ilic. 2023. Uniform pricing vs pay as bid in 100%-renewables electricity markets: A game-theoretical analysis. In Proceedings of the 14th ACM International Conference on Future Energy Systems. 236–241

    2023

  46. [46]

    Tian Zhou, Hao He, Shengjun Pan, Niklas Karlsson, Bharatbhushan Shetty, Brendan Kitts, Djordje Gligorijevic, San Gultekin, Tingyu Mao, Junwei Pan, et al

  47. [49]

    InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining

    An efficient deep distribution network for bid shading in first-price auctions. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 3996–4004. 9 A Appendix A.1 Pseudo-code The offline training, online inference and numerical search compo- nents of our HOB framework are presented in Algorithm 3, 1, and 2, respectively. Alg...

  48. [50]

    Since𝜆> 0,𝑉≥ 0, and𝜋∈[ 0, 1), it follows that𝑒𝜆𝑉≥ 1and1 −𝜋≤ 1

    At the other boundary,𝑥=𝑉 , we haveℎ(𝑉)=( 1−𝜋)[ 1+𝜆(𝑉− 𝑉)]−𝑒 𝜆𝑉 =( 1−𝜋)−𝑒 𝜆𝑉 . Since𝜆> 0,𝑉≥ 0, and𝜋∈[ 0, 1), it follows that𝑒𝜆𝑉≥ 1and1 −𝜋≤ 1. Thus,ℎ(𝑉)≤ 0. The equality holds only in the trivial case where 𝑉= 0and 𝜋= 0, but the conditionℎ( 0)> 0implies 𝑉> 0. Therefore,ℎ(𝑉)< 0. Sinceℎ(𝑥) is continuous and strictly decreasing on[0,𝑉] withℎ( 0)> 0and ℎ(𝑉)< 0...

  49. [51]

    This implies that 𝑔′(𝑥)< 0for 𝑥∈( 0,𝑉] and 𝑔′(0) ≤0

    Since ℎ(𝑥) is strictly decreasing, for any 𝑥∈ (0,𝑉] , we have ℎ(𝑥)<ℎ( 0)≤ 0. This implies that 𝑔′(𝑥)< 0for 𝑥∈( 0,𝑉] and 𝑔′(0) ≤0. Thus, 𝑔(𝑥) is a strictly decreasing function on [0,𝑉] (or non-increasing ifℎ( 0)= 0). The unique maximum is therefore achieved at the left boundary,𝑥∗ = 0. This behavior also satisfies the definition of strict unimodality, as t...

  50. [1995]

    Auctioning the airwaves: the contest for broadband PCS spectrum.Journal of Economics & Management Strategy4, 2 (1995), 345–373

    1995

  51. [2021]

    Automated bidding and budget optimization for performance advertising campaigns.A vailable at SSRN 3913039(2021)

    2021

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.