What Hidden Bids Let Sellers Learn About Auction Reserves

An auction’s minimum price does more than determine whether an item sells: it also hides what lower-valued buyers would have offered. New research examines how sellers can choose reserves when historical records leave those values unknown.

A seller looking through past auction bids faces a blind spot: people whose values fell below the minimum acceptable price may never have submitted a bid. More records collected at the same prices can reveal how often this happened, but not what those missing values were.

In the November 2026 arXiv preprint Offline Learning of Reserve Prices from Censored Bids, Lavanya Marla, Yunzong Xu and Boxuan Zhou study what sellers can still learn. Their results separate uncertainty that additional historical data can reduce from ambiguity that persists even with unlimited records. The findings are presented as preprint research, not established peer-reviewed conclusions.

The information below the minimum price

The paper considers second-price auctions for one item. A reserve is the lowest acceptable price. With one qualifying bidder, the winner pays the reserve; with two or more, the winner pays the second-highest bid.

The model assumes at least two risk-neutral bidders whose private values are independent draws from a common, absolutely continuous distribution with finite mean. Historical reserves and participation counts are fixed before values are drawn. Crucially, the seller knows how many bidders faced each reserve, including those who submitted no bid.

Call the lowest historical reserve the boundary. At the population level, records reveal the distribution of values at and above that boundary and the fraction below it. Different distributions of the hidden lower values can nevertheless produce identical records while favoring different reserves. This is an offline problem: the seller cannot collect new observations by deliberately trying lower prices.

When the best reserve is knowable

The main results assume a monotone hazard rate, or MHR. Here, the hazard describes how quickly bidders cease to qualify as the price rises, relative to those still qualifying. MHR requires that rate to be nondecreasing. It is an assumption about values, not something censoring automatically establishes.

Under MHR, the authors characterize two situations in which all compatible distributions share an optimal reserve.

First, if the hazard immediately above the boundary is sufficiently low, revenue has not begun declining there. The hidden values cannot make a lower reserve preferable to the best choice in the observed range.

Second, the fraction of hidden bidders may reach the largest level MHR permits given the boundary hazard. That equality forces a constant hazard below the boundary, determining the hidden distribution completely.

Outside these cases, compatible distributions disagree about the optimum. Additional records at the same reserves cannot resolve that disagreement.

Choosing a defensible compromise

The authors address this ambiguity using minimax revenue regret. For each possible hidden distribution, regret compares the chosen reserve’s expected revenue with that distribution’s own best expected revenue. The rule minimizes the largest such loss.

In the nonidentifiable MHR setting, they prove that the unique minimizing reserve depends on just two observable quantities, once the target bidder count and historical boundary are fixed: the censored fraction and the hazard immediately above the boundary.

The reserve balances the worst losses from pricing too high and too low. It lies below the lowest historical reserve and has strictly smaller worst-case regret than simply reusing that reserve. This does not mean it earns more under every individual distribution: robustness and performance under one particular distribution are different objectives.

The authors also show that worst-case hidden distributions need at most two constant hazard levels, making the optimization finite-dimensional.

What finite samples add

The learning guarantees depend on the situation. With N historical bidders facing the lowest reserve, an empirical rule achieves expected worst-case regret of order N^(-1/2) when the optimum lies in the observed range, using a known positive lower bound on the boundary hazard.

When the distribution is fully identified and the optimum lies at or below the boundary, a separate rule using only the censored count achieves order N^(-1).

In the nonidentifiable setting, estimating the two quantities yields expected worst-case regret within order N^(-2/5) of the irreducible population minimum. This requires a twice continuously differentiable conditional bid density just above the boundary. A matching lower bound on a smooth interior class limits uniform improvement, not faster learning on particular distributions.

Evidence and practical limits

Synthetic experiments illustrate protection against incorrectly specified distribution families. Recorded eBay bids provide a less controlled test: on Palm Pilot data, the learned rules beat all three censored parametric fits at two of three reference boundaries, but not the third.

These were retrospective evaluations using simulated three-bidder auctions from held-out records, not a live seller experiment. Winner values required imputation; a separate opening-bid analysis also inferred missing participation counts.

The shape assumption matters. Diagnostics found evidence against MHR for Xbox and Cartier records, and the method performed poorly in one heavily censored Xbox setting. The paper’s guarantees therefore remain conditional on its demand, sampling and participation assumptions—not a general promise that robust reserves outperform fitted models.