The winner's curse in common-value auctions

The winner's curse is the tendency of the winning bidder in a common-value auction to pay more than the item is worth. In a common-value auction the item is worth the same to every bidder, but each bidder holds only a noisy estimate, so the highest estimate tends to win. The Nobel committee's 2020 scientific background explains why winning can be "bad news" about value, and credits Capen and colleagues (1971) with the first empirical evidence, from offshore oil-lease auctions. The same document notes that the curse is common among relatively inexperienced bidders in laboratory experiments.

Capen, Clapp and Campbell on oil leases (1971)

The three authors were engineers at Atlantic Richfield. Their paper in the June 1971 Journal of Petroleum Technology says the company first used a probability model for bidding in 1962 and had the model in its published form by 1965. They wrote that, ignoring the era before 1950, the Gulf of Mexico had paid off at less than "the local credit union."

Their explanation was divergence of estimates. Bid ratios between the highest and lowest serious competitors were commonly 5 or 10 and ranged to as much as 100. From that they drew the central claim: in competitive bidding, "the winner tends to be the player who most overestimates true tract value."

In the 1969 Alaska North Slope sale, the second-high bids summed to $370 million against $900 million for the winning bids, so the runner-up bid on average 41 percent as much as the winner. For 26 percent of tracts, the second-high bidder could have quadrupled its bid and still lost. A 50-percent increase would not have won 77 percent of tracts. Their model usually advised lower bids as competition increases, to guard against "the winner's curse."

The jar-of-coins experiment (1983)

Bazerman and Samuelson ran the first laboratory test, reported in the December 1983 Journal of Conflict Resolution. The subjects were 419 master of business administration (MBA) students in 12 microeconomics classes at Boston University. Each class bid in four sealed-bid auctions for jars of pennies, nickels, or paper clips. Unknown to the subjects, every jar was worth $8.00. Subjects also gave an estimate and a 90 percent confidence range, a $2.00 prize went to the closest estimate, and no feedback came until the end. Bidders were told their group size, from 4 to 26, for 48 auctions in all.

The average estimate was $5.13, $2.87 below the true value. Across the 48 auctions, the mean of the average winning bids was $10.01, an average loss of $2.01 to the winner. In 23 of the 48 auctions the average winning bid exceeded $9.00. Uncertainty drove the effect more than group size. In estimates the authors called only suggestive, a 20 percent rise in uncertainty raised the average winning bid 43 percent, while a 20 percent rise in bidders raised it 5 percent. The authors concluded that subjects bid on unconditional estimates and ignored what winning implies.

Kagel and Levin's controlled auctions (1986)

Kagel and Levin's 2001 survey of their experiments describes the design. The common value is drawn at random from a known range, each bidder gets a private signal within a known distance of it, and the high bidder earns the value minus the bid. Bidders start with cash, go bankrupt at zero, and see all bids and the true value after every auction. Results are compared with a risk-neutral Nash equilibrium (RNNE) benchmark. In the first nine auctions with inexperienced bidders, profits averaged -$2.57 against a predicted $1.90, and 40 percent of subjects went bankrupt.

The December 1986 article studied moderately experienced bidders. The abstract reports that equilibrium models fit auctions with 3 to 4 bidders, while auctions with 6 to 7 bidders produced more aggressive bidding and negative profits. Public information raised seller revenue without a winner's curse and reduced it with one. The 2001 survey gives the profits as $4.32 per auction in small groups against an RNNE prediction of $7.48, and -$0.54 in large groups against $4.82.

Thaler's 1988 Anomalies column

Thaler's column, the first Anomalies article in the Journal of Economic Perspectives, gave two versions of the curse. In one, the winning bid exceeds the value. In the other, the value is below the winner's estimate, so the winner is disappointed even with a profit. Thaler noted that the curse cannot occur if all bidders are rational, so evidence of it is an anomaly.

Among Thaler's field studies, Mead and colleagues found an average present-value loss of $192,128 per lease across 1,223 Gulf of Mexico leases issued from 1954 to 1969, at a 12.5 percent discount rate. Hendricks, Porter and Boudreau, using other assumptions, found the same sales profitable, yet 12 of 18 frequent bidders would have earned more by bidding less. In a study by Weiner, Bazerman and Carroll, only 5 of 69 MBA students learned to bid $1 or less on a related takeover problem over 20 trials with feedback.

Thaler's figures differ from two other accounts. For Kagel and Levin, Thaler gives small-group profits as 65.1 percent of the RNNE prediction and large-group losses of $0.88 per period against a predicted $4.68 profit, where the authors' 2001 survey gives the figures above. For the North Slope sale, Thaler says the winning bid exceeded the second by a factor of at least 2 on 77 percent of tracts, while the 1971 paper says only that a 50-percent increase by the second-high bidder would not have won them.

What to check

• Is the value common to all bidders, and does the agent condition its estimate on winning?
• Does the agent shade bids more as rivals are added, or less, as Kagel and Levin's bidders did in groups of 6 to 7?
• Is performance measured against an equilibrium benchmark such as the RNNE, or only against breaking even?
• Did repeated auctions with feedback reduce overbidding, as they did for only 5 of 69 students in one study?
• Does a cited figure come from the original paper or from a later summary, and do the two agree?

Azcuna's CALIBER puts AI agents in practice markets with play money and a news feed whose truth is known, and measures from their trades whether they are gullible, follow the crowd, overpay, or can be steered. It was proposed to DARPA under SBIR Phase I topic DPA26BZ06-DV026, and the proposal was submitted on October 21, 2026.

Sources

1. 1Competitive Bidding in High-Risk Situations · Capen, Clapp and Campbell, Journal of Petroleum Technology, June 1971 (Princeton University course copy)
2. 2I Won the Auction but Don't Want the Prize · Bazerman and Samuelson, Journal of Conflict Resolution 27(4), December 1983 (Princeton University course copy)
3. 3The Winner's Curse and Public Information in Common Value Auctions · Kagel and Levin, American Economic Review 76(5), December 1986 (RePEc record and abstract)
4. 4Common Value Auctions and the Winner's Curse: Lessons from the Economics Laboratory · Kagel and Levin, Ohio State University, October 2001
5. 5Anomalies: The Winner's Curse · Richard H. Thaler, Journal of Economic Perspectives 2(1), Winter 1988
6. 6Scientific Background on the Sveriges Riksbank Prize in Economic Sciences in Memory of Alfred Nobel 2020: Improvements to auction theory and inventions of new auction formats · Committee for the Prize in Economic Sciences in Memory of Alfred Nobel, October 12, 2020

Where this shows up in our work

• CALIBER, the market testbed for AI agents measures from an AI agent's trades whether it is gullible, herds, overpays, or can be steered.
• LEDGER, auditable Army acquisition decisions turns a trade study into a package where every number cites its source and AI drafts but never decides.