Part 1 of a 9-part series simulating a stock market traded by language-model agents. Build the call-auction clearing engine and benchmark it against zero-intelligence controls before any model is involved.

The first part of a series building a discrete-time stock market simulator populated by language-model agents: the call-auction clearing engine, the budget constraint, and zero-intelligence controls benchmarked against Cont's stylized facts.

Agentic Market Simulator, Part 1: The Engine and the Environment

With one clearing event per simulated day, there is nothing for a continuous double auction to be continuous about, since queue priority and intraday matching only matter if participants can act at different moments within the step. A uniform-price call auction is the honest mechanism for that time resolution, and it is what real opening and closing auctions use.

Why use a call auction instead of a continuous limit order book?

Price discovery is not the interesting output. Random agents produce a reasonable price but a terrible allocation, and they cannot hold a view, remember being wrong, change strategy after a drawdown, or react to news in language, which is what the rest of the series studies.

If zero-intelligence agents already track the fundamental value, why use language models at all?

No. Efficiency is realized surplus over available surplus, and an agent that buys above its own valuation contributes negative surplus. The unconstrained zero-intelligence regime does this constantly, so its losses accumulate to many times the surplus that was ever available to capture.

Agentic Systems

Agentic Market Simulator, Part 1: The Engine and the Environment | SR Cookbooks