In a sealed-bid auction for an asset of uncertain common value, an offshore oil tract, say, whose true worth is the same for every bidder but known to none, each firm must submit a bid based on its own private estimate. Because these estimates are noisy, some overshoot the true value and some undershoot it. The crucial observation is that the firm whose estimate is highest is also, on average, the firm that has most overestimated the asset; and since the highest bid generally wins, the winner is disproportionately likely to be the bidder who has erred most on the optimistic side. A firm that bids its honest estimate therefore wins chiefly in those cases where its estimate was too high, and thus systematically overpays. This is the 'winner's curse.' The naïve diagnosis is that bidders are simply too optimistic and should form more accurate estimates. But the curse afflicts even bidders whose estimates are unbiased, correct on average across all possible draws, because conditioning on the event of winning selects the high tail of the estimate distribution. The rational response is not to revise one's estimate of the asset but to revise one's interpretation of what winning means: a bid should be shaded below one's raw estimate by an amount that reflects how much the act of winning itself signals overestimation. That shading should be larger when estimates are noisier and when more rivals are bidding, since each additional competitor makes the winning estimate more extreme. A bidder who fails to shade is not irrational about the asset's value in the abstract; the error lies in neglecting the statistical information embedded in the very fact of having outbid everyone else.