Continuous clearing auction.
An agent-based model of a continuous clearing auction (CCA) for fair-launch token issuance. Six heterogeneous agent types bid into a uniform-price clearing engine across 24 periods; the simulator stress-tests how supply schedules, sentiment regimes, and information leakage shape the clearing trajectory — and who ultimately captures surplus.
Continuous clearing auctions release tokens incrementally over 24 periods. At each step, agents observe a public signal about fundamental value, update beliefs, and may submit a single bid (price cap + budget). A uniform-price clearing engine sets the period’s price; the cumulative auction terminates with one final clearing pass. The design question: can a CCA price an asset close to its fundamental value while remaining robust to whale extraction and adversarial information regimes?
We instantiate six agent archetypes — early believers who anchor on private valuations, informed traders who weight the public signal heavily, momentum traders who chase observed price drift, option-value optimizers who delay entry to learn from earlier clearing prices, noise traders who bid stochastically, and whales who split a large budget across tranches. Each type has its own belief-update rule, entry condition, and bid-cap policy. The full agent specification is in technical/simcluster/cca_abm/agents.py.
TRY IT: AUCTION SIMULATOR
Move the sliders to set sentiment, demand intensity, and information leakage. The clearing-price trajectory and per-agent surplus update in real time — interpolated from a 125-point Monte Carlo grid of the continuous clearing auction.
Sentiment shifts the initial belief distribution up or down — high sentiment pulls bid caps above fundamental, low sentiment compresses them below. The clearing price tends to under-shoot V at low sentiment and over-shoot at high sentiment, but the rate at which it reverts toward fundamental depends on how quickly informed traders accumulate.
Demand intensity is the mean per-bidder budget. Higher budgets thicken the order book and pull the clearing price up; at very low budgets the auction starves and clears at the price floor for several periods before any sustained bidding appears.
Info leakage controls how fast the public signal converges on fundamental value. At high leakage, informed traders dominate early surplus; at low leakage, momentum and early believers carry the price. Whales benefit from moderate leakage — enough signal to know when to enter, not enough for the rest of the book to price them out.
The surplus bars are signed: positive bars mean that agent type captured surplus on average; negative bars mean they overpaid. The fastest way to find an interesting regime is to drag sentiment low while keeping demand high — the model surfaces the configurations where retail noise traders pay the most, and where the protocol’s schedule can be tuned to redirect that surplus.
The widget interpolates a precomputed 5×5×5 Monte Carlo grid over (sentiment, demand, leakage). Each grid point averages three runs of a 24-period auction with the full agent population (12 early, 10 informed, 15 momentum, 6 option-value, 25 noise, 3 whales) drawing budgets from a truncated normal. The supply schedule is uniform; the price-bound box is [$0.10, $50]. Trilinear interpolation over the grid gives a smooth control surface in the browser with no server round-trips.
The full Python implementation — including the bisection-based uniform-price clearing engine, the supply-schedule constructors (uniform, front-loaded, back-loaded, bell curve), and the agent belief updates — is in technical/simcluster/cca_abm/. Re-running the sweep with a denser grid or different agent mix is a single CLI invocation.