Should a Bidder Act, Abstain, or Explore? Separating Censoring from Noise in Selective Bidding

Many automated buyers only see what happens below their own offer. An ad-buying system that loses an auction learns only that the clearing price was higher than its bid. A retailer that sells out learns only that demand exceeded stock. Models trained on such data can describe the market well below the offers that were actually made and know almost nothing above them. Current bid-shading and demand models still produce a confident answer for every request, and current methods that let models abstain assume the true label is always eventually observed. This proposal builds a rule that, for each request, chooses one of three actions. The system acts when the expected outcome is pinned down. It abstains when the market is inherently unpredictable for that context. It explores, bidding or stocking higher on purpose, when the uncertainty exists only because its own past behavior hid the relevant data. Each decision comes with an explanation of which kind of uncertainty drove it.

The opportunity is that censoring-induced uncertainty can be computed exactly. Above the highest offer ever made in a context, the win probability is not estimated badly. It is not identified at all, and its possible range follows directly from the data. Statistical noise is a different quantity with a different remedy. Bidding higher removes the first and does nothing about the second. Abstaining avoids the second and makes the first permanent. Any single confidence score that mixes the two sends the system toward the wrong action. Our claim is that keeping them apart improves surplus at controlled risk and produces explanations that are both stable and checkable.

Core Claim

In right-censored decision problems, total uncertainty about the payoff of a candidate action splits into two parts. The first is an identification width: the range of payoffs consistent with every market price distribution that agrees with the observed wins and losses. The second is a sampling width: ordinary finite-sample error inside the identified region. Write $F(b\mid x)$ for the probability that the market price $z$ is at most bid $b$ in context $x$. Write $\bar b(x)$ for the highest bid with observed outcomes near $x$. Then for $b > \bar b(x)$ the data constrain only

$$ \hat F(\bar b(x)\mid x) - s(x) \;\le\; F(b\mid x) \;\le\; 1, $$

where $s(x)$ is the sampling error of the estimated win rate at $\bar b(x)$. The interval has width $1-\hat F(\bar b\mid x)$ from identification plus roughly $s(x)$ from sampling. The sampling part shrinks like $1/\sqrt{n}$ in the number of nearby auctions, as any binomial proportion does. The identification part does not shrink with more data collected under the same bidding policy.

Parametric censored models, such as the log-normal Deep Distribution Network that An Efficient Deep Distribution Network for Bid Shading in First-Price Auctions uses for production bid shading, fill this gap by extrapolating the distributional shape. Their predictive variance therefore understates exactly the uncertainty that matters most. We test whether a three-way rule (act, abstain, explore) built on this split, with selection-time risk control, beats abstention driven by total uncertainty.

Research Program

The work has three parts, and each reuses an existing technique for a new purpose.

The first part is a censoring-aware selection rule with risk control. We take the conformal selection idea from Act or Defer: Error-Controlled Decision Policies for EHR Foundation Models. That paper turns calibration scores into p-values and applies Benjamini–Hochberg control, a procedure that bounds the expected fraction of false positives among the cases chosen for action. In its setting this bounds errors on the cases acted on. It assumes fully observed labels.

In our setting, the bad outcome at a candidate bid $b'$ is a negative or below-target surplus. That outcome is observed for a calibration auction in two cases: the auction was won, or it was lost at a bid of at least $b'$. It is censored otherwise. We therefore reweight calibration scores by the inverse probability of being uncensored at $b'$. This follows the weighting that Conformalized Survival Analysis for General Right-Censored Data uses to obtain valid lower bounds under general right-censoring. That paper gives coverage without any option to abstain. We give abstention with censoring-corrected guarantees. The same paper warns that these weights become unstable under extreme censoring. The explore branch is our response: once weights exceed a cap, the request is sent to exploration rather than treated as calibratable.

The second part is the decomposition used as an explanation. For each request we report the two widths and the action they imply. We then explain each width at the feature level. For identification width, the explanation is the subset of context features whose neighborhoods have the lowest maximum bid. For sampling width, it is a standard attribution of the variance of a bootstrap ensemble.

Here we take the measurement protocol of Explanation Multiplicity in SHAP: Characterization and Assessment seriously. That paper shows that SHAP rankings churn across random seeds even when predictions are confident. We report Top-$k$ Jaccard distance and Rank-Biased Overlap of our explanations across seeds. We do not report only magnitude intervals of the kind produced by Approximating Score-based Explanation Techniques Using Conformal Regression, because magnitude intervals leave rank instability hidden. We expect the identification component to be far more stable, since it is a deterministic function of logged bids rather than of model fitting.

The third part is transfer to demand censoring. The same machinery applies to ordering decisions in which sales equal the minimum of demand and stock. An order above any past stock level corresponds to a bid above $\bar b$. We use this transfer as a secondary test of generality rather than the main result.

Baselines

The naive reference point is forced action. The censored log-normal DDN, trained with the Tobit-style likelihood (density for wins, survival function for losses), shades every request without ever abstaining. This is the deployed industrial practice. An Efficient Deep Distribution Network for Bid Shading in First-Price Auctions reports a 9.65% offline surplus lift and a 14.3% online surplus lift over point-estimate shading on Verizon Media logs. Those logs are not public, so we reimplement DDN on iPinYou. Below it we also include complete-case training, which discards losses. Bid-aware Gradient Descent for Unbiased Learning with Censored Data in Display Advertising shows that complete-case training underperforms survival-corrected models. It reports a 2.97% AUC lift and a 9.30% eCPC reduction online for the survival-corrected models. We rerun complete-case training only as a sanity floor.

The strongest existing competitor is total-uncertainty abstention. The base predictor is the same DDN. A conformal lower bound on the market price is calibrated with the Fused procedure of Conformalized Survival Analysis for General Right-Censored Data, which reaches its 90% coverage target on NACD, GBSG, METABRIC, SUPPORT and Churn. We add Benjamini–Hochberg selection in the style of STRATCP. The system abstains whenever the selection test fails.

This competitor is strong because it has everything we have except the split. It handles censoring correctly, it controls selected risk, and it can abstain. We give it the same base model, the same calibration split, the same hyperparameter search budget over its thresholds, and the same exploration budget. An exploration budget it spends at random among abstained requests is a fair comparison, because it has no signal for where to explore.

For the market setting, no reported result was found on Lacuna for selective or abstaining bidding on iPinYou. The nearest reported numbers are MINMAX's 7.42% surplus lift and 22.75% profit lift on iPinYou, from Simultaneous Optimization of Bid Shading and Internal Auction for Demand-Side Platforms. That method needs observed clearing prices, so it is reported only in the uncensored oracle condition, where it serves as an upper reference.

Two ablations isolate the key ingredient. In the first, our rule keeps the identification bound but merges both widths into a single abstention score. In the second, the identification bound is replaced by the DDN's parametric extrapolation. Beating the competitor means three things together. The first is higher cumulative surplus at matched selected-risk level $\alpha$. The second is selected risk held at or below $\alpha$, which the competitor must also meet. The third is lower identification width after the same exploration budget. A surplus gain that disappears under the first ablation would mean the split adds nothing.

Why not just…?

Why not just abstain whenever the DDN's predicted variance is high? The parametric model reports low variance precisely above $\bar b$, where it is extrapolating its assumed shape. It would therefore act confidently where it knows least. The second ablation tests this directly. If parametric variance turns out to track ground-truth error above $\bar b$ as well as the identification bound does, this branch of the proposal fails, and we will say so.

Why not just use the fraction of lost auctions near $x$ as the explanation and the trigger for exploration? This is the most serious threat. The censoring fraction is cheap and correlated with identification width. It differs in a way that matters, though. A context can be heavily censored at bids far above the surplus-optimal bid, in which case censoring is irrelevant to the decision. Or it can be lightly censored at exactly the bids that matter. Our width is evaluated at the candidate bid, and the censoring fraction is not. We run the censoring-fraction heuristic as a full competitor for both exploration targeting and explanation. If it matches our method in surplus and in ranking contexts by how much revealing their true prices would help, we will conclude that the simple heuristic suffices.

Why not just explain abstentions with CLUE-style counterfactuals? Getting a CLUE: A Method for Explaining Uncertainty Estimates searches a generative model's latent space for the smallest input change that removes predictive uncertainty. It improved human anticipation of model certainty to 82.22%, against 52.78% for sensitivity maps on tabular data. However, the change that removes censoring uncertainty is a change to the bidder's past actions, not to the input. Moving the input to a well-observed region answers a question nobody asked. CLUE is also only as faithful as its generative model, a limitation that paper itself notes. We include CLUE on the DDN's entropy as an explanation baseline.

Proposed Experiments

The main testbed is iPinYou, from Real-Time Bidding Benchmarking with iPinYou Dataset. It contains 64.75M bid records across 9 advertisers, with paying prices for won impressions. Following the semi-synthetic protocol of the bid-aware gradient descent work, we impose censoring by replaying a lower historical bidding policy. This hides prices that the full log actually contains. The design gives ground truth for every counterfactual bid, which is the decisive feature. We vary the replay policy to sweep censoring rates from moderate to above 95%, since the conformal survival papers flag that regime as unstable. TukMob serves as a second ad dataset. For demand, we apply the same replay to the NYC bike-sharing and supermarket series studied in Diffusion-aware Censored Gaussian Processes for Demand Modelling, imposing simulated stock caps.

The metrics fall into four groups. For outcomes, we report cumulative surplus per unit budget and selected risk against target $\alpha$ in $\{0.05, 0.10, 0.20\}$. For exploration, we report identification width remaining after a fixed exploration budget, along with surplus regret relative to the uncensored oracle. For explanation stability, we report Top-$k$ Jaccard distance and Rank-Biased Overlap across ten seeds. For the extreme-censoring regime, we report coverage.

The decisive check uses the hidden ground truth. For contexts labeled "explore," revealing the hidden prices should cut payoff error substantially. For contexts labeled "abstain," revealing them should barely help. We measure this as the gap in error reduction between the two groups, against the same gap when contexts are sorted by total width and by censoring fraction.

Three outcomes would falsify the proposal. The first is that our labels do not separate the two groups better than the censoring-fraction heuristic does. The second is that selected risk exceeds $\alpha$ once inverse-censoring weights are capped. The third is that, at matched risk, the surplus gain over total-uncertainty abstention is within seed variance.