股票市场仿真系统SMSS — A research platform for limit order book studies

The trader model running in this exhibit is the Main ABM— a 10,000-agent boundedly-rational population built in the spirit of Goettler, Parlour & Rajan, extended with continuous-time Poisson event flow, stochastic-approximation belief learning, and behavioral elements (disposition effect, stop-loss).

This is one model, not the only model. Any other trader population — different decision rule, different scale, reinforcement learning, behavioral experiment, human-in-the-loop — connects through the same protocol. What follows demonstrates what kind of research questions this platform can address, using this particular model as the running example.

Computational Equilibrium — Belief Convergence

The curves below show the Weighted Absolute Deviation (WAD) of agent belief distributions between successive measurement intervals. As WAD falls below the convergence threshold, the population's beliefs — their estimates of fill probability at each price level (μ_b for buy, μ_s for sell) and expected round-trip return (Tvalue) — have stabilized.

The learning is shared.Every agent's observed fill or non-fill updates a common belief table indexed by (system state, action), which all agents query when deciding their next action. This is computational equilibrium emerging from interaction, not designed in, and not solved analytically.

▶ The platform in research mode · Exchange Engine (left) and the Main ABM Trader Monitor (right) running in parallel

Two independent processes, one shared protocol. The exchange runs a continuous auction at ~67K orders/sec; the trader pool runs the Main ABM's 10,000 agents, sharing a common belief table indexed by (system state, action), updated by every fill any agent observes. The Belief Convergence panel (center-right) shows the WAD of the population's μ_s beliefs declining past the threshold — convergence emerging from interaction, not designed in.

μ_b WAD evolution · 12× timelapse · 372 testing rounds

μ_b — Buy-side fill belief

WAD declining from ~0.024 toward the convergence threshold over 306 testing rounds.

μ_s — Sell-side fill belief

Symmetric convergence on the sell side — different scale, same qualitative dynamic.

The Learning, Live — Main ABM Trader Monitor

The WAD curves above compress belief convergence into a single scalar. The clip below is the multi-panel raw viewof that same process: belief tables (μ_b / μ_s / Tvalue) updating with every fill and non-fill any agent observes; individual agents' net worth and holdings evolving through their trading; order flow intensity and distribution spreading across the screen simultaneously. After watching it, “10,000 agents learning through shared interaction” stops being an abstract phrase.

▶ Main ABM Trader Monitor · 45-second live clip showing the learning process (belief tables + individual agent tracking + order flow evolving in parallel)

Emergent Belief Structure

The belief distributions, plotted across 11 limit-order-book price levels (B5 … A5 around mid-price), reveal what agents have learned: the probability of buy-side and sell-side execution at each level. The shape of the learned belief surface emerges from market interaction; the action space and state representation are designed, the values they take are not.

μ_b profile across 11 LOB levels

Population-average buy-fill probability declines smoothly from deep ask (A5, ~1.0) to deep bid (B5, ~0.3) — the learned shape of the order book, not a designed prior.

μ_s profile across 11 LOB levels

Sell-fill probabilities form a complementary structure — what 10,000 agents have jointly learned about both sides of the market.

Inspectable at the Individual Level

Macro-level aggregates emerge from concrete trajectories of individual agents. Any agent in the population can be tracked: net worth, P&L, holdings, and trade history are recorded over the full simulation. Aggregate findings remain auditable down to a single trader — when 10,000-agent emergent dynamics produce a stylized fact, the scientific question is which subset of behaviors causes it, and individual-level inspection answers that.

Five concurrently tracked agents — net-worth trajectories, holdings, recent trades. Drawn from a live simulation of 10,000 agents.

Validated against real markets

Five canonical stylized facts of equity markets — leverage effect, volatility clustering, long-memory volatility, order flow autocorrelation, and excess kurtosis — emerge from agent interaction in the Main ABM, not as designed features.

The Main ABM is one application; the platform supports many more. Below are research and teaching scenarios that the same engine — with different trader populations or configurations — can serve.