QuoteLab

When do tighter quotes earn more, and when do informed traders erase the edge?

QuoteLab is an event-driven research laboratory for synthetic single-asset market making. It compares a fixed-spread market maker with inventory skew, volatility widening, and their combination. Identical exogenous market tapes make policy comparisons paired; every session closes inventory at a cost and reconciles cash to an exact P&L decomposition.

Delivered study: 12,580 sessions, six main market regimes, 30 development seeds and 40 held-out seeds per regime, six base half-spreads, and separate toxicity/refresh sweeps. The core has 33 automated tests. These are synthetic experiments, not a real-market backtest.

Start here

1. Read reports/QuoteLab_Research_Memo.pdf for the question, results, and limitations.
2. Open reports/QuoteLab_Dashboard.html in any modern browser. This saved explorer needs no server, network, account, or additional packages.
3. Follow the commands below to change assumptions and rerun the study.
4. Work through docs/INTERVIEW_GUIDE.md and audit one trade in a CSV ledger.

Run locally

Requires Python 3.10 or newer. From the extracted quotelab directory:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
python -m unittest discover -s tests -v
python -m quotelab dashboard

On Windows use python -m venv .venv and .venv\Scripts\Activate.ps1 instead of source. Then open http://127.0.0.1:8765. The dashboard can compare all four policies on a new shared tape while changing the spread, risk aversion, volatility buffer, inventory limit, fee, or quote-refresh interval. Only the single-path tab changes when you run new settings; the evidence tabs retain the bundled research dataset.

One session, including a path and trade ledger:

python -m quotelab simulate --scenario toxic --policy adaptive --seed 20000
python -m quotelab simulate --scenario jump_shock --policy fixed --refresh 8

Reproduce the complete experiments (about one minute on the development machine; runtime depends on hardware):

python -m quotelab research --workers 4
python -m pip install -e '.[report]'
python scripts/build_report.py

--workers 1 is the serial default and produces the same ordered rows. Versions used in the delivered run are in results/manifest.json and requirements-reproduction.txt. Numerical libraries can change random sampling or last-digit floating-point results between versions. The manifest records seed ranges, parameters, versions, session counts, and the maximum accounting residual. The PDF and figures are built from the CSV results, so changing the study also changes the report's data.

What the study found

At the reference base half-spread of $0.04, inventory skew reduced average absolute inventory in the toxic regime from 15.23 to 1.44 shares, but its mean net P&L was -$27.28. Adding volatility widening produced mean P&L of $30.17 and mean maximum drawdown of $2.82 in that regime.

The interpretation matters: widening changes the actual spread. After each policy's base width was selected on development seeds, the combined policy's toxic-regime P&L advantage over the tuned fixed baseline had a paired 95% interval containing zero. Its inventory and drawdown remained much lower. The combined policy also underperformed the tuned fixed baseline in mean P&L in some regimes.

Refresh stress provides a failure case: in toxic flow, the combined policy's mean P&L fell from $29.67 with one-step refresh to -$55.37 with eight-step refresh. A reactive volatility estimate cannot guarantee protection against fresh jumps or stale quotes.

Design

Component	Responsibility
config.py	Units, validation, policy defaults, six synthetic regimes
market.py	Arithmetic jump-diffusion price path; ordinary/informed arrivals; immutable shared tape
strategy.py	Causal quotes using current observed mid, own inventory, and EWMA variance
engine.py	Ordered fills, per-fill risk limits, paid liquidation, markouts, accounting identity
metrics.py	Seed-level bootstrap intervals and paired policy effects
experiment.py	Disjoint selection/evaluation seeds, grids, feature ablations, sensitivity sweeps
server.py	Loopback-only dashboard and bounded simulation endpoint
scripts/build_report.py	Reproducible figures, memo, and offline interactive explorer

Model in one minute

At step t, a policy observes mid[t], its inventory, and past returns. It submits or refreshes a bid and ask. The efficient price then moves to mid[t+1], and an independently generated sequence of potential orders arrives.

Ordinary traders decide using the prior public midpoint mid[t] plus an exponential willingness-to-pay depth. Informed traders see the new efficient price and trade only when a standing quote offers them more than their minimum edge. Quotes auto-refill one share at a time at unchanged prices until the next refresh; the inventory cap is checked at every fill. No queue position, competing makers, price impact, or network cancellation dynamics are modeled.

The dealer's signed fill quantity is dq (+1 buy, -1 sell), price is p, and the quote was formed at reference midpoint m_ref:

quote edge       = dq * (m_ref - p)
fill repricing   = dq * (m_exec - m_ref)
inventory P&L    = q_before_step * (m_exec - m_previous)
net P&L          = sum(quote edge + fill repricing + inventory P&L)
                   - execution fees - terminal liquidation costs

Quote edge includes inventory skew and ticks. Fill repricing can be positive or negative; it is not universally a loss. informed_execution_pnl separately isolates the net execution edge of filled informed orders, which is negative by construction. Neither repricing nor markouts are subtracted from P&L a second time.

Research protocol

• Main grid: 6 regimes x 70 seeds x 6 base half-spreads x 4 policies = 10,080 sessions.
• Toxicity sweep: 4 informed-arrival fractions x 25 seeds x 4 widths x 4 policies = 1,600.
• Refresh sweep: 3 stress regimes x 25 seeds x 3 intervals x 4 policies = 900.
• Development seeds: 10000-10029. Evaluation seeds: 20000-20039.
• Sensitivity seeds: 30000-30024. A common seed is a common exogenous path within each regime.
• Fixed reference base half-spread: $0.04. Additional plots show the entire grid.
• Selection: one width per policy, chosen only from development results, with equal regime weights and objective mean P&L - 0.25 * mean maximum drawdown.
• Inference: 3,000 percentile bootstrap resamples across complete seed-level sessions. Paired effects first subtract policies on the same seed. Trades within a session are correlated and are never treated as independent observations.
• Intervals measure Monte Carlo uncertainty under the specified model. Exploratory grid comparisons have no multiple-testing correction. Seed holdout is not a real data time split, and it cannot establish market profitability.

Audit the artifacts

results/runs.csv contains one metrics row for every session. path_* files contain sampled state and cumulative accounting components for the illustrative seed 20000; trades_* files contain every passive fill. Closing costs and the flattened final position are in the path and metrics, not the passive trade ledger. The final two path rows share a timestamp: the first is before close, the second after liquidation.

results/reference_summary.csv, paired_deltas.csv, selected_parameters.csv, and selected_heldout_summary.csv separate reference effects from development- selected comparisons. toxicity_summary.csv and refresh_summary.csv hold the stress results. demo_paths.json preserves the full reference simulations.

Scope and next experiments

The original contribution here is the combined experiment and its diagnostics: causal information boundaries, exogenous shared tapes, exact accounting, controlled feature ablations, and failure-case analysis. The inventory center is inspired by Avellaneda-Stoikov; the implementation is an explicitly documented heuristic with EWMA variance, passive constraints, a volatility buffer, jumps, and informed flow. It does not solve an HJB or reproduce that paper's experiment.

The most valuable next step is a price-time-priority order book with competing quotes and finite refill/cancel latency. A subsequent public-data replay would require trade/quote alignment, conservative queue assumptions, and a chronological calibration/evaluation split. More simulator features alone do not validate an edge.

Primary references:

• Avellaneda and Stoikov, 2008, High-frequency trading in a limit order book
• Jane Street, Probability & Markets