A market-microstructure research platform that reconstructs NASDAQ limit order books from LOBSTER data and replicates canonical order-flow results — order-flow imbalance, queue-imbalance prediction, trade-sign long memory, price impact — with honest statistics (HAC errors, purged walk-forward CV, FDR control, confidence intervals on every effect). Its distinguishing feature is that the statistics pipeline was verified against a ground-truth-injectable simulator — it recovers planted signals and stays silent on noise — before it touched real data.
This project reports effect sizes with confidence intervals under multiple-testing control, and nothing more. There are no PnL, strategy, tradability, or forecasting claims anywhere — the measured effects are real but small and cost-dominated, and the research note says so plainly. Descriptive microstructure research, not investment advice.
In one line: reconstruction is provably exact against an external oracle; the statistics pipeline provably recovers planted signals and stays silent on noise; and the replicated effects are real, small, and cost-dominated — every one reported with a confidence interval under FDR control.
| Metric | Value |
|---|---|
| LOBSTER events processed | 2,110,860 (AAPL · AMZN · GOOG · INTC · MSFT, 2012-06-21, level 10) |
| Reconstruction correctness oracle | 100.0000% exact per-event transition invariant — 0 violations / 1,673,089 transitions¹ |
| Pipeline verification | injected β recovered within 0.05%, 97.5% CI coverage (40 seeds), power rises to 100%; placebo FPR 2.5% |
| OFI → contemporaneous price | HAC slope = 5.8e-9 (p ≈ 1e-192); R² = 0.150 [0.120, 0.186] as a descriptive fit (Cont–Kukanov–Stoikov) |
| OFI → forward price | positive & FDR-significant at every horizon (direct out-of-sample coef, per-symbol purged walk-forward CV); coef is return-per-unit-OFI, so small by construction, rising 1.3e-10 (h=1) → ~2.1e-9 (h=100) |
| Queue imbalance → next move | AUC = 0.637 [0.617, 0.658]; 0.91–0.92 in large-tick names² (Gould–Bonart) |
| Trade-sign long memory | ρ decays 0.764 → ~0 over lags 1–200 (Bouchaud) |
| Price impact | concave, log-log exponent 0.026 [0.005, 0.046] |
| Registered study family | 17 tests, pre-registered; 15 FDR-significant at α = 0.05 |
| Test coverage / determinism | ≥90% (CI-gated), ~97% on CI, 160+ tests; every statistic & figure reproduces bit-for-bit |
¹ The free LOBSTER level-10 sample is level-restricted, so a full-depth message-only book match is
structurally impossible; the exact per-event price-level transition invariant is the valid
correctness oracle and it holds at 100%. See docs/DESIGN.md §D1.1.
² Large-tick names (INTC ~$28, MSFT ~$31 — spread pinned to one tick) vs high-price names (AAPL ~$586, GOOG ~$580): AUC 0.907/0.923 vs 0.556/0.578, with AMZN ~$224 mid-regime at 0.614 — the canonical tick-size signature.
flowchart TD
M[LOBSTER messages] --> R[independent book reconstructor]
R --> D[transition-invariant differential<br/>vs LOBSTER reference book — the oracle]
B[reference book] --> F[point-in-time features<br/>OFI · queue imbalance · tick-rule sign · depth · spread · realized vol]
F --> E[estimators<br/>OLS+HAC · logistic · IC · AUC · Bartlett ACF · block bootstrap]
E -->|purged & embargoed walk-forward CV| FAM[pre-registered 17-test family]
FAM --> FDR[Benjamini-Hochberg FDR] --> OUT[effect sizes + confidence intervals]
SIM[ground-truth simulator] -.->|recovery · placebo · leakage · goldens<br/>must pass BEFORE any real-data study — the iron rule| E
Contracts (src/schema.py, src/interfaces.py) and the pre-registration
(config/registry.py: every alpha, tolerance, horizon, and the study family, each justified) were
frozen before the machinery was built. Every methodological choice is logged in
docs/DESIGN.md; the full results write-up is in
docs/RESEARCH_NOTE.md.
- Reconstruction oracle — the exact per-event price-level transition invariant (submit,
cancel, delete, execution) vs LOBSTER's own book: 100%, 0 violations / 1,673,089
transitions, tolerance 0. The full-depth message-only match rate is reported honestly as a
data-limitation figure, never target-matched (
DESIGN §D1.1). - Synthetic recovery — a ground-truth-injectable simulator; the pipeline recovers the injected β within 0.05% with 97.5% CI coverage and power rising to 100%.
- Placebo nulls — β = 0 yields a 2.5% false-positive rate, at/below nominal.
- Leakage detection — constructed leaks (label overlap, future-timestamped features,
un-embargoed boundaries) are each caught (
tests/test_cv_leakage.py,test_integration_g1.py). - Estimator goldens — HAC vs statsmodels, AUC vs scikit-learn, IC vs scipy, Bartlett ACF vs statsmodels — all to 1e-8 on synthetic data.
- Multiple testing — Benjamini-Hochberg across the whole 17-test family
(
src/multi/fdr.py, golden-matched to statsmodels; empirical FDR ≤ α verified). - Determinism — everything seeded from one master seed;
runner,g1_gate, and every figure regenerate byte-identically. The alpha/tolerance registry holds every threshold with a written justification, never widened to force a pass.
python3.12 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]" # or: make install
make verify # ruff + format-check + CI test subset + ≥90% coverage gate
# or run the pieces directly:
python -m pytest -m "not slow" # verification suite (CI subset), ~97% coverage
# reproduce everything (needs the LOBSTER sample):
make data # download the sample (not committed)
make gate # print the machinery-gate evidence (python -m src.studies.g1_gate)
make studies # rerun the 17-test family -> data/fixtures/*.csv
make figures # regenerate docs/figures/*.png (deterministic)
python -m pytest # full suite incl. slow real-data checks- Verified-machinery-first. The iron rule: no real-data study runs until the pipeline has recovered planted signals, stayed silent on noise, and caught constructed leaks. "I tested my testing."
- Point-in-time by construction. Every feature declares its information timestamp; look-ahead is structurally impossible and enforced by leakage tests.
- Honest claims. Effect sizes with CIs, FDR-controlled, zero tradability language; surprising and null results (the small return-per-unit-OFI magnitudes, the lag-100 and lag-200 non-rejections) are kept and discussed, never dropped.
- Reproducible. One documented command per published statistic; seeded determinism throughout.
docs/RESEARCH_NOTE.md— the ~10-page write-up: verification of the machinery first, then results with CIs, the literature comparison, the tick-size discussion, the transaction-cost caveat, and limitations.docs/DESIGN.md— the methodology log: every non-trivial choice with its rationale and literature anchors.docs/ENV.md— pinned toolchain and how to reproduce the environment.
Uses LOBSTER (lobsterdata.com) limit-order-book sample data, which
reconstructs NASDAQ order books from TotalView-ITCH. Raw files are not redistributed here —
they are downloaded at build time by scripts/get_data.sh under LOBSTER's academic-use terms;
only small derived fixtures are committed. Please cite LOBSTER if you use the data. Provenance of
the exact sample used is in docs/DESIGN.md §D0.1.
MIT (source code only — see LICENSE, including the note on data).