VI

The market model

VNQuant has no real data. Prices form when about 300 simulated investors and the players meet in the order book under Vietnamese exchange rules. This page describes the model, publishes its calibration and lists every known limitation.

The exchange

Each simulated day runs an ATO call auction, 48 continuous-trading steps, then an ATC auction. The book uses price-time priority; auctions pick the volume-maximising price. LO, MP, ATO and ATC orders; HOSE ±7% band, ticks and 100-share lots; cash held when a buy is placed; shares settle at T+2; 0.15% fee each way and 0.1% sell tax.

Players and simulated investors go through the same order gate. Cash and shares are integers; conservation is tested after every order.

Companies and fundamental value

12 fictional companies in 6 sectors, with 4-letter tickers so they never match real ones. Each has a fundamental value V driven by a factor model: market factor, sector factor and its own noise, scaled by the company's betas.

All shocks are scaled by a shared, slowly varying "information intensity" (the mixture-of-distributions hypothesis, Clark 1973). This is the main source of volatility clustering.

News

Company, sector and macro stories arrive at random (a Poisson process) and move V up or down. News is published during the session; prices don't jump at once but adjust as simulated investors update their views.

Simulated investors

Each stock has 1 market maker (two-sided quotes, Ho–Stoll inventory skew), 8 value investors (noisy, lagged estimates of V; buy below it, sell above it), 6 trend followers (buy after rises over their window, more active when the market is agitated) and 10 noise traders.

No agent is programmed to produce fat tails or volatility clustering. They emerge from the interaction.

Calibration results

Default scenario, measured with npm run calibrate. Ranges are 95% confidence intervals across seeds. Measured on 2026-11-18 · 10 × 250 seeds × sessions

Stylized factTargetResultMet
Fat tails (excess kurtosis of daily returns)> 01.48 [1.05; 1.91]Met
Returns nearly uncorrelated (lag 1)|ρ| < 0.10.00 [-0.04; 0.05]Met
Volatility clustering (|r| autocorrelation lag 1 → lag 10)positive, decaying0.23 → 0.08Met
Volume correlated with |r|> 00.43 [0.36; 0.49]Met
Daily volatilityreference2.19% [1.92; 2.47]—
Share of closes at ceiling/floorreference1.83% [0.69; 2.98]—

Known limitations

  • The unfilled part of an MP order is cancelled; the real rule turns it into a limit order.
  • No odd lots, margin, short selling, dividends or splits.
  • Settled shares are merged and held at average cost.
  • An account may trade against its own order.
  • The final call-auction tie-breaker (higher price) is the project's choice; exchange rules need checking against current regulation after the KRX system launch.
  • No financial statements for the fictional companies yet.
  • In the lab, the same seed takes different paths under different rules as soon as behaviour changes; comparisons rely on distributions across many seeds, not single paths.
  • This is a simplified model; results are for learning and asking questions, not for forecasting real markets.
Research lab →