What is a second-price (Vickrey) sealed-bid auction?

Every bidder privately submits one sealed bid without seeing anyone else's. The bidder with the highest bid wins the item or ad slot, but instead of paying their own bid (as in a first-price auction) they pay the amount of the second-highest bid. William Vickrey introduced this mechanism in 1961 and won the Nobel Memorial Prize in Economics in 1996 largely for this and related work in auction theory. Real ad-exchange auctions historically used variants of this rule because of the strong incentive properties it creates for bidders.

Why is bidding your true value a dominant strategy in a second-price auction?

In a second-price auction your bid only ever affects two things: whether you win, and never how much you pay if you win (that is fixed by the second-highest bid, which you do not control). Bidding above your value can only ever make you win auctions you would rather have lost -- cases where the price you must pay exceeds your value, producing negative utility. Bidding below your value can only ever make you lose auctions you would rather have won -- cases where the second-highest bid is below your value, so winning would have produced positive utility you gave up. Bidding exactly your value never creates either of these bad outcomes, no matter what any other bidder does. Because this holds for every possible configuration of competitors' bids, truthful bidding is a dominant strategy -- it does not depend on guessing what anyone else will bid.

What actually happens if a bidder shades their bid up or down, and does this simulation prove it or just assert it?

This simulation runs a genuine computed auction every round: fresh random valuations are drawn from the same distribution for every strategy group, sealed bids are computed from each bidder's actual strategy, bids are sorted, and the winner pays the real second-highest bid. Utility (value won minus price paid, zero for losers) is summed and divided by the number of bidder-rounds for each strategy group live, so the on-screen averages are the actual output of thousands of real simulated auctions, not scripted numbers. Because all three groups draw from an identical valuation distribution, the law of large numbers guarantees their average valuations converge to the same value as rounds accumulate -- so any persistent gap in average utility you see between the truthful, shade-up and shade-down lines is attributable to the strategy itself, and it reliably shows the truthful line settling above the other two.

Why did real online ad auctions historically favor second-price-style mechanisms?

Google's original AdWords auction (and many ad exchanges since) used a generalized second-price (GSP) mechanism precisely because asking advertisers to bid their true value simplifies the bidding problem enormously: advertisers do not need to run game-theoretic simulations of competitors to decide what to bid, they can simply report their genuine value per click or per impression and let the mechanism handle efficient allocation and pricing. This lowers the cognitive and strategic burden on participants and, in the single-item Vickrey case, provably maximizes total economic efficiency (the item always goes to whoever values it most) while remaining strategy-proof for each bidder individually.

Does increasing the number of competing bidders change whether truthful bidding is optimal?

No -- dominant-strategy incentive compatibility means truthful bidding remains a bidder's best response regardless of how many competitors exist or how they behave, because the proof never referenced the number or behavior of competitors, only the mechanical fact that your bid cannot influence the price you pay if you win. Adding bidders in this simulation changes how often the truthful group wins (more competition means lower win rates for everyone and a second-highest bid that sits closer to the winner's true value) and narrows everyone's average utility toward zero, but it never flips the ranking: truthful bidding still weakly dominates shading in either direction.

Does widening the valuation spread or increasing the shading percentage change the conclusion?

Widening the valuation spread increases the variance of outcomes round to round but does not change the underlying dominant-strategy logic, since the argument for truthful bidding holds value-by-value, not just on average. Increasing the shade-up or shade-down percentage sliders makes the non-truthful strategies fail more severely -- heavier shade-up bidders win more often but overpay by more when they do (sometimes turning a positive-value win into a net loss), while heavier shade-down bidders lose auctions they could have profitably won more often. Turning either shading slider down toward zero makes those groups' performance converge back toward the truthful group's, which is exactly what the theory predicts: shading is only ever neutral or harmful, never beneficial, in a second-price auction.

首页▸人工智能与机器学习▸广告拍卖竞价者 — 实时第二价格维克里拍卖

📢 广告拍卖竞价者 — 实时第二价格维克里拍卖

实时观看一场真实的第二价格(维克里)密封竞价广告拍卖,展示随着出价值和竞争者数量变化,诚实出价为何是每个广告主的占优策略。

人工智能与机器学习3D中等60 FPS

关于广告拍卖竞价者

如今几乎每一次程序化广告展示的售卖都是通过密封竞价拍卖完成的,而第二价格(维克里)规则正是这一市场机制设计的理论核心:出价最高者赢得广告位,但支付的只是第二高出价的金额,而不是自己的出价。这一单一设计选择带来了一个非凡的结果——由诺贝尔奖得主威廉·维克里于1961年证明:出你的真实私人估值成为一种占优策略,意味着无论任何竞争者出价多少,这都是你的最佳选择。本模拟不只是断言这一结果,它会真正计算出来:每一轮,一个可配置的模拟竞价者群体——有的诚实出价,有的按固定百分比上调出价,有的下调出价——都会从相同的分布中抽取新的私人估值,提交密封出价,并解出一场真实的拍卖(排序、选出赢家、按第二价格收费)。每种策略群体的效用(赢得的价值减去支付的价格)会实时累加。

由于每个群体都从同一分布中抽取估值,大数定律保证随着轮次累积,它们的平均估值会收敛到同一个数值——这意味着实时平均效用图表中任何持续存在的差距,都可归因于出价策略本身,而不是谁恰好更看重这个广告位的运气。观察这个三维柱状场:每个竞价者都用一个半透明的轮廓框表示其真实价值,一个实心彩色框表示其密封出价,因此压价情况可以直接看作两者之间的差距;一个金色平面标出实时的第二价格支付水平,一个金色圆环加冕本轮的赢家。调整每种策略的竞价者数量、估值分布范围和压价百分比,观察这种效应增强或减弱——但随着足够多轮次的累积,你会看到实时图表中诚实出价(橙色)线始终稳定在上调压价(红色)和下调压价(蓝色)线之上。

常见问题

什么是第二价格(维克里)密封竞价拍卖?

每个竞价者私下提交一个密封出价,看不到其他任何人的出价。出价最高的竞价者赢得物品或广告位,但不像第一价格拍卖那样支付自己的出价,而是支付第二高出价的金额。威廉·维克里于1961年提出了这一机制,并主要因这项研究及相关的拍卖理论工作于1996年获得诺贝尔经济学纪念奖。真实的广告交易所拍卖历史上曾使用这一规则的变体,因为它为竞价者创造了强有力的激励特性。

为什么在第二价格拍卖中出真实价值是一种占优策略?

在第二价格拍卖中,你的出价只影响两件事:你是否获胜,而绝不会影响你获胜时要支付多少(那是由第二高出价决定的,而这不受你控制)。出价高于你的真实价值,只会让你赢得那些你本该宁愿输掉的拍卖——即你必须支付的价格超过你的价值,产生负效用的情形。出价低于你的真实价值,只会让你输掉那些你本该赢得的拍卖——即第二高出价低于你的价值,本可获得正效用却主动放弃了。恰好出你的真实价值,无论其他任何竞价者出价多少,都不会产生这两种糟糕结果中的任何一种。由于这对竞争者出价的每一种可能配置都成立,诚实出价是一种占优策略——它不依赖于猜测其他任何人会出多少。

如果竞价者上调或下调出价,实际会发生什么?这个模拟是证明了这一点,还是只是断言?

本模拟每一轮都运行一场真正被计算出来的拍卖:每个策略群体都从相同的分布中抽取新的随机估值,根据每个竞价者的实际策略计算出密封出价,对出价进行排序,赢家支付真实的第二高出价。效用(赢得的价值减去支付的价格,输家为零)被累加,并针对每个策略群体实时除以竞价者轮次数,因此屏幕上的平均值是数千场真实模拟拍卖的实际输出,而不是脚本编写的数字。由于所有三个群体都从相同的估值分布中抽取样本,大数定律保证它们的平均估值会随着轮次累积收敛到同一个值——因此你在诚实、上调压价和下调压价这三条线之间看到的任何持续差距,都可归因于策略本身,而它可靠地显示诚实线最终稳定在另外两条线之上。

为什么真实的在线广告拍卖历史上偏爱第二价格类机制?

谷歌最初的 AdWords 拍卖(以及此后的许多广告交易所)使用了广义第二价格(GSP)机制,正是因为要求广告主出真实价值极大简化了竞价问题:广告主不需要对竞争者进行博弈论模拟来决定出价,他们只需报告每次点击或每次展示的真实价值,让机制来处理高效的分配和定价。这降低了参与者的认知和策略负担,并且在单一物品的维克里拍卖情形下,可证明能在保持对每个竞价者策略无懈可击的同时,最大化总体经济效率(物品总是归属于最看重它的人)。

增加竞争竞价者的数量会改变诚实出价是否为最优的结论吗?

不会——占优策略激励相容性意味着,无论存在多少竞争者或他们如何行动,诚实出价始终是竞价者的最佳应对,因为该证明从未涉及竞争者的数量或行为,只涉及一个机械事实:你获胜时要支付的价格不受你出价的影响。在本模拟中增加竞价者会改变诚实群体获胜的频率(更多竞争意味着每个人的获胜率都更低,第二高出价也更接近获胜者的真实价值),并使每个人的平均效用都趋近于零,但它永远不会颠倒排序:诚实出价在任何一个方向上仍然弱占优于压价。

拓宽估值分布范围或增加压价百分比会改变这一结论吗?

拓宽估值分布范围会增加每一轮结果的方差,但不会改变底层的占优策略逻辑,因为诚实出价的论证是逐个价值成立的,而不仅仅是平均成立。增大上调或下调压价百分比滑块会使非诚实策略失败得更严重——压价更重的上调型竞价者会更频繁获胜,但获胜时支付得也更多(有时会把一次正价值的胜利变成净亏损),而压价更重的下调型竞价者会更频繁地输掉那些他们本可以盈利赢得的拍卖。将任一压价滑块调低到接近零,会使这些群体的表现趋近于诚实群体,这正是理论所预测的:在第二价格拍卖中,压价永远只会是中性或有害的,绝不会带来好处。

⚙ Under the hood

Each round, every bidder draws a fresh private valuation from a shared distribution, computes a sealed bid from its strategy (truthful, shade-up %, or shade-down %), and a real auction resolves: highest bid wins, price paid is the second-highest bid. Utility accrues per strategy group into a running average shown live and charted over rounds.

Vickrey AuctionGame TheoryDominant StrategyMechanism DesignMonte Carlo

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install