Teaching with Auction Lab

A free, open-source tool for showing an auction mechanism actually running — bids arriving, a winner being chosen, a pricing rule firing — one step at a time, with the rule that fired written out in English beside the numbers it produced.

What it is for

Auction theory is taught with a payment rule on a slide and a worked example beside it. The gap students fall into is between the two: they can restate the rule and still not see why a second-price auction makes honesty safe, because the argument is about what would have happened had they bid differently. This runs both, side by side, with their own numbers in it.

How to use it in a lecture

• Build the case live. Type the bidders from your own lecture notes, run it, and step through. Nothing is precomputed and there is no wrong setup.
• Send the exact case to the class. Every setup is a URL. Press Copy link and the auction you just built reopens for anybody who has it — in a slide, a syllabus, or a problem set.
• Show the deviation, do not assert it. Change one bid and re-run. The claim "you cannot do better by lying" becomes something a student watches fail to happen.
• Run it over rounds. The same bidders can meet repeatedly and adjust as they learn, with budgets and a value distribution, which is where pacing and convergence stop being footnotes.

Mechanisms

• Second-price (Vickrey) — Highest bid wins and pays the highest losing bid. Truthful bidding is dominant.
• First-price sealed bid — Highest bid wins and pays its own bid, which rewards shading below your value.
• All-pay auction — Highest bid wins, but every eligible bidder pays their bid — losing bids are sunk.
• English (ascending clock) — Price rises until one bidder is left; equivalent to a second-price auction.
• Dutch (descending clock) — Price falls until someone claims the item; equivalent to a first-price auction.
• Greedy combinatorial auction — Bids are accepted in descending order, skipping conflicts; each winner pays their own bid. Practical, not truthful.
• VCG combinatorial auction — The best allocation is solved for exactly, then each winner is charged the welfare the others lose by their presence. Truthful.
• Generalized second-price (GSP) — Slots go to the highest bids; each winner pays the next bidder's bid per click. Not truthful.
• VCG position auction — The same slots to the same bidders, but each winner pays the externality they impose. Truthful.

Worked examples

• Honesty is safe — Three bidders all name their true value. The winner pays the runner-up's bid, not their own — which is why nobody gains by lying.
• Why everyone shades — Pay your own bid and honesty earns you nothing. Watch four bidders best-respond round after round, and see where the bids stop falling.
• Losers pay too — Everyone pays what they bid and only one of them wins anything. Three of these four bidders end the auction underwater.
• A reserve too high — The seller refuses anything under 80 and the room tops out at 72. Nobody wins, nobody pays, and the item everyone wanted goes nowhere.
• Ads: four slots, one ladder — Search ads sell position, not the page. Higher slots get more clicks, so what each bidder pays depends on who is standing below them.
• Bundles don't add up — These bidders want combinations, not items — a pair is worth more than its halves. Greedy grabs the best-looking bid first and leaves money behind.
• The budget runs out — Twenty impressions, real budgets, fresh values each round. One bidder spends until they are broke; the other paces itself and is still bidding at the end.
• Where does it settle? — Not what happens at these bids — what happens at every other bid too. The reply curves say whether anyone still has a reason to move.
• No stable bid at all — Every profile in an all-pay auction has somebody who wishes they had bid differently. That is not the absence of an equilibrium — it is one made of dice, and here is what the dice look like.

What it costs and what it collects

Nothing, and nothing. There is no account, no login, and nothing that identifies a visitor is stored. It is MIT-licensed and runs on the Python standard library with a dependency-free front end, so a department can host it themselves from the repository if that is easier than pointing at this one.

If a mechanism you teach is missing, open an issue. Adding one is a single decorated generator, and the interface builds itself from the registry.

Open the walkthrough