Loading SMSS exhibit…
Loading SMSS exhibit…
首页/Exhibition Hall 01 — Stock Market Simulation System
An independent limit order book exchange engine that accepts any trader model — algorithmic, learning, or human — through a standard protocol, with browser-based real-time monitoring from any machine.
Not a simulator. Infrastructure built to support a 5–10 year research program in market microstructure.
▶ LIVE · EXCHANGE MONITOR · CONTINUOUS AUCTION · 60K ORDERS/SEC · 10,000 AGENTS TRADING
② Why Build From Scratch
The doctoral research required 10,000 heterogeneous agents trading in real time, each independently learning through stochastic approximation, while maintaining strict price-time priority matching, supporting the complete A-share call auction mechanism, and producing tick-level market data for statistical analysis. General-purpose agent-based modeling frameworks — Mesa, NetLogo, MASON, ABIDES — were designed for breadth across simulation domains rather than the engine-level performance and pluggable trader interface this question demanded. No existing open platform met all of these requirements simultaneously.
▶ 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 ~85,000 orders/sec; the trader pool runs the Main ABM's 10,000 agents, each updating μb, μs, and Tvalue beliefs from observed fills. The Belief Convergence panel (center-right) shows the WAD of the population's μs beliefs declining past the threshold — convergence emerging from interaction, not designed in.
Theoretical Foundation
The agent-based model demonstrated here builds on the computational equilibrium framework of Goettler, Parlour, and Rajan (Journal of Financial Economics, 2005; Review of Financial Studies, 2009) — one of the foundational models for understanding strategic behavior in limit order book markets.
This implementation extends the framework to 10,000 agents with continuous-time Poisson event flow and Pakes–McGuire stochastic approximation belief learning, enabling the study of equilibrium convergence dynamics at a scale previously inaccessible.
A Python prototype (v1) produced the empirical results for the doctoral thesis. The current system (v2) re-implements the validated approach with production-grade engineering — transforming a working research instrument into a sustainable platform.
③ Proof
The simulation reproduces market dynamics consistent with empirically observed patterns. Beyond statistical validation, the system provides direct visual evidence of computational equilibrium — the state where 10,000 agents' independently learned beliefs stabilize through repeated market interaction.
Stylized Facts vs Real Markets
| Metric | SMSS | Real |
|---|---|---|
| Leverage Effect | 1.88 | 2.0 (DAX) |
| Long Memory d(|r|) | 0.504 | 0.48–0.55 (DAX) |
| GARCH Persistence | 0.980 | 0.989 (DAX) |
| Order Flow Hurst | ≈0.7 | ≈0.7 (Empirical) |
| Fat Tails (Kurtosis) | 3.67 | ~10 (DAX) |
Five canonical stylized facts of equity markets — leverage effect, volatility clustering, long-memory volatility, order flow autocorrelation, and excess kurtosis — emerge endogenously from agent interaction, not as designed features.
Computational Equilibrium — Belief Convergence
The curves below show the Weighted Absolute Deviation (WAD) of agent belief distributions between successive measurement intervals. As the WAD falls below the convergence threshold, the beliefs of 10,000 agents — their learned estimates of fill probability and expected profit — have stabilized. This is computational equilibrium: a state where no agent finds reason to significantly revise its trading strategy, emerging purely from repeated interaction in the order book.
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.
Emergent Belief Structure
The order-book-aligned belief distributions reveal what agents have learned: the probability of buy-side and sell-side execution at each price level relative to the mid-price. Structure emerges as agents discover the market's microstructure through experience, with no exogenous priors.

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.

μ_s profile across 11 LOB levels
Sell-fill probabilities form a complementary structure — what agents jointly know about both sides of the market.
④ Platform
SMSS separates the exchange engine from trader models so that new research questions require only a new model — not a new system. The decoupling runs deep: the engine, the order protocol, the trader pool, and the monitoring dashboards each evolve independently.
Complete A-Share Trading Day — Native Implementation
Unlike general-purpose simulation frameworks that model only continuous auction, SMSS natively implements the complete trading day of the Chinese A-share market — directly applicable to research on price discovery during auctions and informed trading across market phases.
| Time | Phase |
|---|---|
| 9:15–9:20 | Opening Call Auction |
| 9:20–9:25 | Opening Call Auction |
| 9:25–9:30 | Opening Match + Indicative Price |
| 9:30–11:30 | Continuous Auction (AM) |
| 13:00–14:57 | Continuous Auction (PM) |
| 14:57–15:00 | Closing Call Auction |
Two Independent Components
The exchange and the trader pool are separate processes that meet at the order protocol. Each opens with a self-describing landing page, exposing the engine for inspection before any orders flow.

Exchange · :8080
A continuous double auction with strict price-time priority matching. Receives orders over ZMQ; publishes the order book and trades over WebSocket to any browser on the network.

Main ABM · :8081 · one of many possible models
The Main ABM (Goettler–Parlour–Rajan, extended to 10,000 agents) is the trader model in this exhibit. Any other model — different decision rule, different population size, reinforcement learning, human-in-the-loop — connects through the same protocol without touching the engine.
Capability Surface
What the platform offers, independent of any particular trader model:
Exchange Engine
Price-time priority limit order book, A-share six-phase trading day, asynchronous Parquet persistence. 460,000+ orders/sec peak; 15% utilization at 10K agents leaves 85% headroom for richer experiments.
Standard Order Protocol
ZMQ + msgpack on TCP. Any trader process — algorithmic, reinforcement learning, behavioral, or human-driven via a thin adapter — connects through the same six-symbol interface. The Main ABM (10,000 agents) shown in this exhibit is one example; the platform is not bound to it.
Browser-Anywhere Monitoring
Two real-time dashboards (Exchange :8080, Trader :8081) over WebSocket. Multiple viewers — researchers, students, collaborators — can observe the same simulation concurrently from any host on the network.
Tick-Level Persistence
Trade intents, LOB snapshots, and trade records persist to columnar Parquet. Configuration, seeds, and git hashes accompany every run, enabling exact replay and downstream econometric analysis.
⑤ Impact
SMSS provides a controlled laboratory for computational finance research — and a ready-made experimental platform for teaching and collaboration. The same engine serves all three.
Research
Run large agent populations to study how individual learning aggregates into market-level equilibrium, with real-time visualization of the convergence process and full belief-trajectory data for analysis.
Plug different trader models into the same exchange to study how market properties change with agent heterogeneity, strategy composition, or information structure — the engine is shared, the model varies.
Test alternative auction rules, matching algorithms, or regulatory constraints in a fully instrumented environment with tick-level data capture and Parquet output for downstream econometrics.
Teaching
Students write their own trader models against the Python SDK and deploy them into a live exchange — turning microstructure theory into hands-on experimentation in a graded environment.
Real-time candlesticks, order book dynamics, and belief evolution displayed on screen — market microstructure made visible, with the lecturer free to pause, modify the schedule, or pose what-if questions live.
The protocol natively supports human participants trading via browser alongside AI agents — enabling controlled experiments in behavioral finance with the same exchange used for purely computational research.
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.

Five concurrently tracked agents — net-worth trajectories, holdings, recent trades. Drawn from a live simulation of 10,000 agents.
Built for the Long Run
The platform currently runs 10,000 agents at 15% engine utilization. The architecture — distributed ZMQ communication, modular trader processes, network-accessible monitoring — supports multi-instrument and multi-machine deployment as research questions demand them.
“Not an endpoint. Infrastructure designed to support the next decade of computational finance research.”
⑥ Try It
A self-contained demo package — engine, trader pool, and both monitoring dashboards. Double-click to launch; no installation, no configuration, no dependencies.
30-second preview · Exchange Monitor with 10,000 agents trading
Includes Exchange Engine, Main ABM (10,000 agents), and dual monitoring dashboards. The package runs locally; nothing is uploaded.
System Requirements