Clock auction simulator

A working ascending clock auction, in the same invented slot-allocation domain as the data platform demo. Nothing here derives from client systems.

Why rebuild rather than sanitise

I have production auction software that I cannot publish. The tempting shortcut is to take it and find-and-replace the client’s names out.

That is a bad idea, and worth saying why. Redaction by substitution leaves comments, variable names, magic numbers, rule thresholds and the shape of the business logic intact. A careful reader reconstructs a surprising amount, and you cannot un-publish it once it is out. Rebuilding the mechanism from the public literature in a different domain takes less effort than auditing a redaction you can never fully trust, and the result demonstrates the same understanding.

The format

Each product is a block of landing slots at an airport in a time window. The auctioneer posts a price per product. Bidders state how many slots they want at that price. Any product with more demand than supply gets a price rise; anything already in balance holds. The auction ends when nothing is oversubscribed.

Two details matter more than they look:

Only oversubscribed products rise. Raising a product that already cleared would break the balance it just found. Getting this wrong produces an auction that never terminates.

Indifference is not a bid. A bidder demands units worth strictly more than the price, not weakly more. Treating indifference as demand silently biases the clearing price upward, and it is the kind of off-by-one that produces plausible, defensible, wrong numbers.

The activity rule is the design

Eligibility is measured in points, and bidders must keep using a set share of theirs each round. Whatever they do not use, they lose permanently.

That single rule is what turns a clock auction from a price-discovery exercise into a strategic one. Without it, bidders sit out the early rounds, let others reveal their valuations, and re-enter late. With it, holding back is a commitment you cannot reverse.

The simulator exposes the rule as a slider precisely because the effect is so large. Set it to 100% and bidders are locked into their opening position. Drop it toward zero and price discovery collapses. It is the most instructive control on the page.

Three bidders behaving differently

• Straightforward demands whatever is profitable at the current price. Truthful, and the benchmark the theory is written against.
• Demand-reducing deliberately bids for less than it wants once a product looks contested, to slow the clock. Individually rational, collectively inefficient. This is the central strategic problem of multi-unit auctions and the reason simultaneous formats attract the criticism they do.¹
• Budget-capped wants more than it can pay for, so it drops units by worst surplus per pound until it fits. Budget constraints distort allocation in ways valuations alone do not predict.

The scoreboard reports revenue and efficiency separately, because they move independently. A design can raise more money while putting slots in the wrong hands, and hiding that behind a single headline number is how bad mechanisms get defended.

Grounding

The simultaneous ascending format and its activity rules come from Milgrom’s account of the FCC design,² with Cramton’s early assessment covering how it behaved once real bidders were let near it.³ The clock variant implemented here, and the reasoning for clock rounds ahead of any package stage, follows Ausubel, Cramton and Milgrom.⁴

Engineering notes

The engine is pure. Given a configuration it produces the same auction every time, with a seeded PRNG for bidder valuations, and it has no clock, no IO and no browser dependency. The interface runs the whole auction up front and then steps through the returned rounds, so the visualisation cannot influence the outcome.

27 tests cover the parts that would otherwise fail quietly: that prices rise monotonically, that balanced products hold their price, that eligibility is genuinely irreversible, that efficiency never exceeds the benchmark, that allocations never exceed supply, and that unsold slots are reported rather than hidden.

One test I got wrong first time is worth recording. I asserted that a product with zero supply would run to the round cap, expecting it never to clear. It clears instantly and correctly, because no bidder can demand a slot that does not exist. The test was wrong, not the engine, and the fix was to test the round cap with a genuinely contested product instead.

Footnotes

1. Ausubel, L. and Cramton, P. (2002). “Demand Reduction and Inefficiency in Multi-Unit Auctions.” University of Maryland working paper. ↩

2. Milgrom, P. (2000). “Putting Auction Theory to Work: The Simultaneous Ascending Auction.” Journal of Political Economy 108(2), 245-272. ↩

3. Cramton, P. (1997). “The FCC Spectrum Auctions: An Early Assessment.” Journal of Economics and Management Strategy 6(3), 431-495. ↩

4. Ausubel, L., Cramton, P. and Milgrom, P. (2006). “The Clock-Proxy Auction: A Practical Combinatorial Auction Design.” In Cramton, Shoham and Steinberg (eds), Combinatorial Auctions, MIT Press. ↩