---
project: anomaly-atlas
document: P001 — The Artifact Frontier, Part I
author: Simon-Pierre Boucher
contact: contact@spboucher.ai
data_source: hfmarketdata.io
created: 2026-08-12
status: published
pub_id: P001
version: 1
title: "The Artifact Frontier: What Survives Honest Testing in Open High-Frequency Market Data?"
subtitle: "Part I — Methods, artifact taxonomy, and train-period results (2000–2016)"
abstract: "We systematically scan open 1-minute-to-daily market data (hfmarketdata.io, sole source) for short-horizon mean-reversion, lead-lag, and calendar anomalies, under a pre-registered protocol: every detector must first pass a synthetic-data gate; every scan runs against measured artifact nulls; hypothesis budgets are declared before testing; and validation layers (multiple-testing correction, transaction costs) are applied in sequence. On the 2000–2016 train split, 62% of the 372 searched rules are naively 'significant' and 18% survive Hansen's SPA — yet the survivors carry physically implausible paper Sharpes, and a deliberately included known artifact (the SPX→SPY 'lead') survives statistical correction unharmed. The cost layer then eliminates essentially everything: the median surviving rule breaks even at 1.1% of one half-spread per trade, and zero intraday rules survive paying the full half-spread. The calendar family — including turn-of-month, the last published survivor — dies against a permuted-calendar null with an 8-test budget. Our principal positive contributions are a measured artifact taxonomy for this dataset and a demonstrated three-layer validation doctrine: artifact nulls, search correction, and costs are independent filters, and no one of them substitutes for another. Out-of-sample confirmation on the untouched validation split is reported in Part II."
---

# The Artifact Frontier — Part I

*Every number in this document regenerates from committed `results.json`
files, and every figure below is rendered live from them by the site. The
[research log](/doc/research/LOG.md) is the audit trail; the
[charter](/doc/CLAUDE.md) is the pre-registered protocol.*

## 1. Question and doctrine

Given only open high-frequency market data — 1-minute to daily bars, no
quotes — which statistical regularities are *real* (reproducible
out-of-sample, robust to artifacts, economically nonzero after costs), and
which are plumbing? The project's doctrine, stated before any data was
touched: **every candidate anomaly is an artifact until proven otherwise;
in-sample results are never findings; negative results are first-class.**

Three design rules operationalize this:

1. **Test the tests.** No detector touches real data before passing a
   synthetic gate: it must find *nothing* in a pure random walk, recover
   planted effects, and flag a pure bid-ask-bounce series as artifact.
   The gate is not ceremonial — it caught two real bugs before they could
   contaminate results (a variance-ratio estimator reading ~1/q on random
   walks, and an end-freshness-only synchronization mask that silently
   attenuated correlations via the Epps mechanism).
2. **Pre-registration.** Universes, time splits (train 2000–2016 /
   validation 2016–2021 / sealed holdout 2022→), hypothesis budgets (22,
   declared in [research_gaps](/doc/research/research_gaps.md)), and every
   experiment's falsification criterion were frozen in writing before the
   corresponding run.
3. **A frozen cache as the reproducibility anchor.** All data flows through
   one client that never silently refetches; every response is indexed in a
   committed data manifest. Two of the seven experiments below ran with
   **zero network requests**, entirely from the frozen cache.

## 2. The data, measured

Experiment A established empirically what the source actually provides:
true 1-minute bars for ~7,700 stocks and ~5,200 ETFs from January 2000
(futures and indices from 2008, FX from 2010, crypto from 2013), plus daily
options chains with quotes and Greeks over 67 quarters. Load-bearing facts
that the documentation does not state: timestamps are US-Eastern wall-clock,
bar-start labeled; **bars exist only where trades occurred** (no zero-volume
placeholders — an illiquid name printed 38 bars in a full session); daily
bars carry the official auction close and consolidated volume, both absent
from the 1-minute series; and dividend-adjusted prices are re-based to the
vendor's build date, so adjusted series are not point-in-time stable.
Responses are hard-capped at 50,000 rows. Full profile:
[data_source_profile](/doc/research/data_source_profile.md).

## 3. The artifact taxonomy, with magnitudes

The project's first deliverable is a catalogue of the mechanisms in *this
dataset* that manufacture fake anomalies — each with detection code and a
measured magnitude ([artifact_taxonomy](/doc/research/artifact_taxonomy.md)).
Experiment B measured the key ones on a pre-specified 42-ticker universe
(Q1 2024, RTH 1-minute):

{{figure:expB_artifact_baselines}}

Highlights: bid-ask bounce alone produces AC1 of −0.23 and VR(30) of 0.55
in the stalest liquidity tercile with zero planted economics; LOCF joins
make SPY spuriously "lead" mid-staleness names (+0.047 at +1 min, Spearman
vs staleness +0.43) while *diluting* the lead of ultra-stale names — the
artifact is non-monotone; and the SPX index print lags SPY by one minute
(+0.065 at 0.965 contemporaneous correlation) — a Fisher (1966) effect,
measured live. The intraday profile (Experiment E) adds the time-of-day
dimension: volatility is U-shaped (6.6 bp at the open, 2.4 midday, 2.9 at
the close) while the effective spread declines monotonically (2.8 → 1.2 bp)
— any "first-30-minutes" return claim fights 2–3× the midday artifact
level.

## 4. The scans (train split only, Level 0 by construction)

**Mean-reversion (Experiment C).** 127 cells (ticker × timeframe ×
sub-period), each tested against a bounce null and FDR-corrected. The
scan's most valuable output was about the null itself: a daily effective
spread combined with pure Roll alternation predicts *impossible* intraday
autocorrelations (−3 to −27), because consecutive intraday closes rarely
flip sides. The corrected, variance-consistent triage — an MA(1) null that
absorbs *all* lag-1 effects — leaves 14 cells of genuine multi-lag
reversion, concentrated in a daily 2008–2015 mega-cap/index family
(XOM excess AC1 −0.13, SPY −0.055, both FDR) and a few 1-minute cells
(JPM −0.24 VR-excess at the 30-minute horizon).

{{figure:expC_reversion_scan}}

**Lead-lag (Experiment D).** 49 pairs, two windows, with the raw-LOCF
versus both-fresh comparison built in — so the non-synchronicity artifact
is *measured*, not consumed. In 2006–2007, minute-scale market→component
diffusion was real on synchronized samples (SPY led every sector ETF by
+0.07..+0.15). By 2014–2015 it had collapsed to ±0.05 — the cleanest decay
measurement of the project. Two structural facts survive synchronization:
a splice-invariant ES↔SPY cross-serial effect (−0.032, identical across
all three futures adjustments), and the SPX→SPY "lead" (+0.132) — which
survives *because synchronizing print times cannot fix a computed index*.

{{figure:expD_leadlag_scan}}

**Calendar (Experiment E).** Eight pre-declared tests (day-of-week ×5,
turn-of-month, pre/post-holiday) on SPY against a within-year
permuted-calendar null with a family-wise max-statistic. **Nothing
survives** (best marginal p = 0.24; family-wise p ≥ 0.93 everywhere).
Turn-of-month — the last survivor in the published literature as of 2006 —
fails and decays inside the train period (+7.9 bp in 2000–2007 → +1.6 bp
in 2008–2015). Both pipeline controls behaved: the Monday effect stayed
dead, and the volatility U-shape was strongly present.

{{figure:expE_calendar_scan}}

## 5. The survival curve: statistical correction is not artifact correction

Experiment F pushed *everything the scans searched* — 372 signed rules —
through the correction battery: naive t-tests, Benjamini–Hochberg FDR,
White's Reality Check and Hansen's SPA over stationary bootstraps, and the
Deflated Sharpe Ratio.

{{figure:expF_multiple_testing}}

The result that matters is not the 18% SPA survival rate — it is *what*
survives: rules with paper Sharpes of 10–31 annualized, physically
implausible, dominated by bounce harvesting (a contrarian rule mechanically
earns −autocov₁ on paper, which is precisely the spread it would pay in
reality). The canary proves the point: the SPX→SPY rule — an artifact we
had already measured twice — **passes SPA comfortably**. Statistical
correction corrects for *search*; it is structurally blind to *mechanism*.

## 6. The cost frontier closes the loop

Experiment G re-priced the double-filtered pool (rules that beat both the
artifact nulls and the search correction; 31 rules) under a declared cost
model: net = gross − κ · (EDGE half-spread) · turnover, κ swept from 0 to 2.

{{figure:expG_cost_frontier}}

The median rule breaks even at **κ\* = 0.011** — it captures about 1% of
one half-spread per trade. Three rules survive κ = 0.1 (all in sparse
names), one survives κ = 0.25 (CKX, an ultra-sparse name with a wide, noisy
spread estimate — the classic profile of an estimation artifact, forwarded
to the validation split with a skeptical prior rather than discarded by
hand), and **zero rules — none — survive paying the full half-spread.**
The pre-registered falsification clause ("the costs-kill story fails if any
intraday rule survives κ = 1") did not trigger.

## 7. What Part I establishes

1. **A three-layer validation doctrine, demonstrated rather than asserted.**
   Artifact nulls, search correction, and transaction costs filter
   *different* failure modes; each layer passed things the next one killed.
2. **A measured artifact taxonomy (T1–T7)** for open bar data, with the
   magnitudes above and neutralization rules, validated on synthetic ground
   truth.
3. **Negative results with teeth**: the calendar family is empty under an
   honest budget; minute-scale lead-lag decayed an order of magnitude
   between 2006 and 2015; and nothing in the searched universe pays for its
   own spread on the train split.
4. **Methodological findings**: spread-based bounce nulls must be
   variance-consistent with the target series; synchronization on print
   times cannot de-artifact computed indices; LOCF distortion of lead-lag
   is non-monotone in staleness.

## 8. Limitations

All Part-I numbers are train-split, in-sample by design — their
out-of-sample fate on the untouched 2016–2021 validation split is Part II.
Bar data carries no quotes: costs are estimated (EDGE), not observed.
Holiday-class calendar tests are low-powered (n = 144). The universe is
U.S.-equity-centric; crypto and FX hypotheses await volume-semantics
verification. Capacity is out of scope.

## 9. Reproducibility

Everything regenerates from the repository: commit `f59e891` (data profile,
client), `9b6beea` (artifact baselines), `c4977f2` (reversion scan),
`a7f66cb` (lead-lag scan), `b228862` (calendar scan), `7a82cc1` (survival
battery), `881399d` (cost frontier). Each experiment's `results.json`
embeds the hardware manifest and the client's instrumentation; the data
manifest (`data_manifest/index.jsonl`) indexes every API response consumed.
Detector gates: `benchmarks/synthetic/` (29 tests at the time of writing).

## References

Key sources (full annotated list with DOIs:
[bibliography](/doc/research/bibliography.md)): Roll (1984); Fisher (1966);
Scholes & Williams (1977); Epps (1979); Lo & MacKinlay (1988, 1990);
Sullivan, Timmermann & White (2001); White (2000); Hansen (2005); Bailey &
López de Prado (2014); Harvey, Liu & Zhu (2016); McLean & Pontiff (2016);
Novy-Marx & Velikov (2016); Chordia, Roll & Subrahmanyam (2005); Ardia,
Guidotti & Kroencke (2024); Chen & Velikov (2022).