When Price-Pushing Meets Noisy Rankings

Online ad platforms rank advertisers by bid multiplied by an estimated click or conversion rate. These estimates are much noisier for new campaigns than for established ones. A separate literature shows that bidders are not always purely self-interested. Some are spiteful: they gain from raising what a rival pays. Platforms also create the same effect on purpose, handing coupons to lower-ranked advertisers so that those advertisers bid up the price the winner pays. The two literatures have never met. Work on spite and coupons assumes the auctioneer sees bids and quality perfectly. Work on noisy selection assumes the candidates are passive boxes that do not bid.

This note argues that the meeting point changes the economics of price-pushing. In a second-price auction with exact scores, a spiteful loser can raise its bid almost for free, because it only pays if it overtakes the winner, and it knows where the winner stands. Once the platform's quality estimate is noisy, that safety margin disappears. Every extra unit of pushing buys a chance of winning an item at a price above one's value, and that chance grows with the variance of one's own estimate. We therefore predict three things. Noise acts as a tax on spite. Platform coupons given to high-variance, cold-start advertisers are wasted or actively harmful. The usual fixes for noisy estimates, such as filtering or shrinking noisy candidates, quietly remove the price pressure that coupons and spite were supplying. We propose to prove this in a stylized model, compute equilibria numerically, and test a variance-aware coupon rule on a public production bidding dataset.

Core Claim

Consider a score-weighted second-price auction. Bidder $i$ bids $b_i$ and has a true quality $q_i$ (its click-through rate). The platform sees only an estimate $\hat q_i = q_i + \epsilon_i$, with noise $\epsilon_i \sim \mathcal N(0,\sigma_i^2)$ whose spread $\sigma_i$ differs across bidders. The platform ranks by score $s_i = b_i \hat q_i$ and charges the winner the second-highest score divided by its own $\hat q$.

Following Spiteful Bidding in Sealed-Bid Auctions, each bidder maximizes a weighted mix of its own payoff and rivals' losses:

$$U_i = (1-\alpha_i)\,u_i - \alpha_i \sum_{j\neq i} u_j .$$

Here $u_i$ is value minus payment if $i$ wins and zero otherwise, and $\alpha_i \in [0,1]$ is $i$'s spite.

Our central claim concerns bidders who expect to lose. For such a bidder, the equilibrium overbid (bid minus value) falls as its own noise $\sigma_i$ rises, and it approaches zero once the noise is large relative to the gap in true scores. The reason is that the chance of an accidental win at an inflated price rises with $\sigma_i$.

The same logic applies to coupons. Coupon Design in Advertising Systems proves that a coupon $c_i$ makes $v_i + c_i$ the dominant bid. When the platform ranks on noisy $\hat q_i$, it has effectively decided that this bidder should push. The expected revenue from a coupon is then the extra price squeezed out of the winner, minus the losses when the pusher's noisy score carries it to the top. That net return is highest for low-variance bidders ranked just below the top. It can turn negative for high-variance ones.

The new object is a coupon or loser-subsidy rule that conditions on estimate variance. The new analysis shows how noise, spite, and variance corrections interact in equilibrium.

How This Sits Relative to Prior Work

Two lines of work motivate the proposal by exposing the gap. The spite papers characterize price-pushing precisely but assume exact scores. Spiteful Bidding in Sealed-Bid Auctions shows that under symmetric spite, second-price auctions out-earn first-price ones. Asymmetric Spite in Auctions shows that this ranking reverses when spite is asymmetric, reporting 0.63 versus 0.49 expected revenue at true spite levels 0.9 and 0.1. Optimal Auctions for Spiteful Bidders derives the revenue-optimal design, which pays subsidies to losers. None of these includes estimation error. We keep their utility models and add a noisy ranking signal.

The noisy-selection papers supply the method on the platform side but have no bidders. Reward Selection with Noisy Observations proves three results. Picking the highest noisy estimate can be $O(n)$ times worse than optimal. Subtracting a fixed multiple of $\sigma_i$ is $\Omega(\log n)$ worse. Discarding high-noise candidates (its IgnoreLarge rule) is constant-factor optimal under a small-tail condition. It explicitly names ad auctions as motivation. Regret Minimization with Noisy Observations gives per-item offsets derived from pricing thresholds. We reuse both as candidate allocation rules. Our question is what happens to bids once bidders anticipate these rules.

Two papers constrain the design. Fairness in Selection Problems with Strategic Candidates is the closest strategic model. It shows that high-noise candidates exert more effort to break through noise. That seems to cut against our "noise taxes pushing" prediction. The two differ because its candidates want to be selected, whereas a price-pusher wants to lose at a high price. We will verify the sign flip directly by switching the objective within one simulator. On Fair Selection in the Presence of Implicit and Differential Variance shows that Bayesian shrinkage suppresses exceptional high-variance candidates. We will measure whether shrinkage also suppresses their price pressure.

Research Program

The first component is analytic. With two bidders, Gaussian noise, and uniform values, the pusher's first-order condition trades the marginal price increase imposed on the winner against the marginal probability of overtaking, weighted by the loss $b_i - v_i$ it would then incur. We will derive the expressed overbid as a function of $(\alpha_i, \sigma_i, \sigma_j)$. The paper will then check whether the Sharma–Sandholm revenue reversal widens or shrinks when the more spiteful bidder is also the noisier one.

The second component is numerical, since closed forms will not extend beyond two bidders. We will compute Bayes–Nash equilibria for $n$ up to 8 by iterating best responses on discretized bid grids, the same fictitious-play approach used by Emelianov et al. We will do this under four platform rules: raw scores, Bayesian-shrunk scores, IgnoreLarge filtering, and OFFSET-style scores.

The third component is the mechanism. The coupon rule assigns $c_i$ as a function of rank position and $\sigma_i$, fit with the Difference-of-Convex surrogate from the coupon paper but with the objective's expectation taken over the estimation noise. The constraint is a fixed coupon budget together with a cap on the rate of negative payments, meaning cases where the platform pays out more than it collects.

Baselines

The naive reference is a plain score-weighted second-price auction on raw noisy estimates, with no coupons and self-interested bidders. It anchors every revenue and misallocation number.

The strongest existing method is the feature-based coupon design of Coupon Design in Advertising Systems. In its synthetic setting it reports a 22.54% revenue gain, against 2.92% for standard reserve prices, with a negative payment ratio below 0.5% on 177M industrial auctions. It is the strongest competitor because it already optimizes engineered price-pushing at scale. We will reimplement it and give it the same coupon budget, the same tuning grid, and the same training auctions as our rule. Crucially, it will also receive $\sigma_i$ as an input feature. Beating it means higher revenue at an equal or lower negative-payment ratio, across three or more seeds. The gap must also grow with the share of high-variance bidders.

A second strong baseline is the pipeline a practitioner would build: shrink or filter the estimates first (Bayesian posterior, IgnoreLarge, or OFFSET), then run the variance-blind coupon optimizer on the corrected scores. Beating it shows that the correction and the coupon must be designed jointly.

For spiteful bidders, the reference is the loser-subsidy mechanism of Optimal Auctions for Spiteful Bidders, evaluated with noisy scores inserted.

The key ablation keeps our full optimizer but removes the noise expectation from the objective. This isolates whether modeling accidental wins is what drives any gain.

Why not just…?

Why not just subtract a variance penalty from each estimate before the auction? Reward Selection with Noisy Observations proves that any fixed penalty is $\Omega(\log n)$ suboptimal even without strategic bidders. With pushers present, a penalty also changes who pushes and how hard. We include penalty and offset rules among the platform rules and report equilibrium revenue for each.

Why not just shrink to the prior and then use existing coupon design? This is the real threat. If shrinkage already neutralizes accidental wins, a variance-blind optimizer on shrunk scores may do as well as a joint design. Shrinkage is known to bury exceptional high-variance candidates (On Fair Selection in the Presence of Implicit and Differential Variance), which suggests it may also bury useful price pressure, but that has not been measured. The two-stage baseline settles it. If it matches our rule within noise, we will report that the contribution is the equilibrium theory rather than a new mechanism.

Why not just drop high-variance advertisers from coupon eligibility, IgnoreLarge-style? This thresholded rule may capture most of the benefit. It is included as a named baseline. We expect it to lose where the small-tail condition fails, or where mid-variance bidders are the best pushers. If it matches us everywhere, we will adopt it as the recommendation.

Proposed Experiments and Decisive Checks

Three benchmarks are used. First, we reproduce the 2-bidder asymmetric-spite testbed and must recover 0.63 and 0.49 at zero noise before adding noise. Second, we use Liu et al.'s 8-advertiser log-normal simulator from Optimal Auction Design with User Coupons in Advertising Systems. Third, we use CriteoPrivateAds. There, per-campaign click estimates and their variance come from bootstrapped logistic regressions (reported rescaled log-likelihood 0.170–0.237), so the privacy noise and cold-start sparsity supply realistic heterogeneous $\sigma_i$. Values on Criteo are simulated, so its results are semi-synthetic.

The metrics are revenue, welfare, the accidental-win rate (pushers who win at a price above their value), and the negative payment ratio.

No reported result was found on Lacuna for any auction that combines spiteful or coupon-induced bidding with noisy quality scores. All comparisons are therefore run here under matched budgets.

The decisive checks are four. First, the computed overbid must fall with $\sigma_i$ for expected losers, and the effort-maximizing objective in the same simulator must reproduce the opposite sign seen by Emelianov et al. Second, under variance-blind coupons, the negative payment ratio and the accidental-win rate must rise with the cold-start share of bidders. Third, our rule must beat both the variance-feature coupon baseline and the shrink-then-coupon baseline at equal budget. Failing the third check means the mechanism claim is dropped.