Independent statistics + market-microstructure research · Apple Silicon · hfmarketdata.io
Which market "anomalies" are real — and which are artifacts?
A systematic, reproducible atlas of statistical regularities in open high-frequency
market data — mean-reversion, lead-lag, calendar effects — where every claim survives (or visibly fails)
out-of-sample testing, multiple-comparison correction, artifact nulls, and realistic transaction costs.
Negative results are first-class findings.
detectable in-sample ≠ reproducible out-of-sample ≠ robust to artifacts ≠ meaningful after costs
A finding only advances one level at a time, and only Level ≥ 1 is ever published.
A large in-sample effect with zero out-of-sample survival is a negative result —
published as one.
Level 0 — in-sample onlyscan output; never published as a finding
Level 1 — corrected & OOSsurvives multiple-testing correction and a clean out-of-sample split, artifact null subtracted
Level 2 — robust+ robust to specification choices, sub-periods, instruments
Level 3 — cost-real & held-out+ economically nonzero after realistic costs, confirmed on the once-touched holdout
The atlas publishes only Level ≥ 1 findings: effects that survive multiple-testing correction
and a clean out-of-sample split with the artifact null subtracted. Nothing has earned that yet —
the framework is being built so that nothing can enter without earning it. The
artifact taxonomy and the
validation methodology are primary deliverables in their own right.
`;
const body = `
The atlas
Confidence-labeled statistical regularities (and non-regularities) in hfmarketdata.io data.
Every entry carries provenance — commit, config, data-manifest hash, hardware — and the exact command that
regenerates it. Negative results are first-class entries.
Formal write-ups of the project's results, versioned in the repository like everything
else. Every number is regenerated from committed results.json files and every figure is rendered
live from them — nothing is hand-typed. Each publication states its evidence level; negative results are
published with the same care as positive ones.
${cards || '
No publications yet.
'}
`;
page(res, { title: "Publications", active: "Publications", body });
});
app.get("/publications/:slug", (req, res) => {
const pub = listPublications().find((p) => p.slug === req.params.slug);
if (!pub) return notFound(res);
let { html, toc } = R.markdownToHtml(pub.content, pub.rel);
// live figures: {{figure:experiment_id}} tokens -> server-rendered SVG
html = html.replace(/
\s*\{\{figure:([\w-]+)\}\}\s*<\/p>|\{\{figure:([\w-]+)\}\}/g,
(m, a, b) => charts.figuresFor(a || b, C) || "");
const d = pub.data;
const body = `
Append-only, newest first — every question, experiment, result,
interpretation, and decision, as required by the charter (§12). ${entries.length} entries.
Every experiment carries a registered hypothesis with an explicit falsification criterion
and its artifact null(s), a benchmark implementation, raw results, and an analysis. Detectors must first pass the
synthetic-series tests — a detector that finds anomalies in a random walk is broken. Negative results are kept.
Every result is reproducible from commit hash + config + data-manifest index + seed +
hardware manifest, and each JSON embeds the manifest of the exact machine that produced it.
Experiment
Run (UTC)
Files
${rows || '
No result runs yet.
'}
`;
page(res, { title: "Results", active: "Results", body });
});
app.get(/^\/results\/(.+)$/, (req, res) => {
const rel = req.params[0].startsWith("results/") ? req.params[0] : "results/" + req.params[0];
const raw = C.readText(rel);
if (raw === null) return notFound(res);
let bodyContent;
if (rel.endsWith(".json")) {
const obj = C.readJson(rel);
bodyContent = `
Core library (src/anomaly_atlas/) — including the single
hf_client every byte of data flows through — tooling, benchmark harness, and experiment
implementations. Every file carries the project's author header and the hfmarketdata.io data-source attribution.
${sections}
`;
page(res, { title: "Code", active: "Code", body });
});
app.get(/^\/file\/(.+)$/, (req, res) => {
const rel = req.params[0];
if (!C.isTextFile(rel)) return notFound(res);
const raw = C.readText(rel);
if (raw === null) return notFound(res);
const name = rel.split("/").pop();
const body = `
`;
page(res, { title: name, active: "Code", body });
});
// ---------------------------------------------------------------- about
app.get("/about", (req, res) => {
const body = `
About this project
anomaly-atlas is an independent research project by
Simon-Pierre Boucher (contact@spboucher.ai) that asks:
given only open high-frequency market data from hfmarketdata.io
(the sole data source), which statistical regularities — mean-reversion, lead-lag relationships,
calendar/seasonal effects — are real, and which are artifacts of bid-ask bounce, stale or
non-synchronized timestamps, survivorship, look-ahead, or microstructure noise?
The methodology is deliberately strict: pre-specified hypotheses and splits, artifact nulls that every
detector must beat, mandatory multiple-testing correction, a holdout touched exactly once, realistic
transaction-cost models, an append-only research log, hardware manifests embedded in every result, and no
result reported from an uncommitted tree. Detectors are validated on synthetic series first — a detector that
finds anomalies in a pure random walk is broken. The full methodology is codified in the
research charter.
What counts as a real result
A regularity reaching Level 3: corrected, out-of-sample, robust, cost-real, confirmed on the once-touched holdout.
A rigorous artifact taxonomy for this dataset — letting others avoid mistaking artifacts for signals.
A clean demonstration that a widely-cited anomaly does NOT survive honest testing here (negative result).
A methodological contribution: a better artifact null or honest test for short-horizon effects on bar data.
A measured survival curve: what fraction of naively-significant effects remains after each layer of rigor.
What this is not
Not a trading system, not trading advice, not a claim of arbitrage or "free money". The honest answer
"nothing survives" is an acceptable — and publishable — outcome. Do not assume Level-3 findings exist;
the research establishes the truth either way.
Platform
Everything runs on a consumer Apple Silicon Mac: DuckDB + parquet on the internal NVMe for out-of-core
columnar queries, NumPy/Accelerate for the math. This site is generated from the repository itself —
no hand-typed numbers.
`;
page(res, { title: "About", active: "About", body });
});
// ---------------------------------------------------------------- comments
function loadComments() {
try {
return JSON.parse(fs.readFileSync(COMMENTS_PATH, "utf8"));
} catch {
return [];
}
}
function saveComment(entry) {
fs.mkdirSync(path.dirname(COMMENTS_PATH), { recursive: true });
const all = loadComments();
all.push(entry);
fs.writeFileSync(COMMENTS_PATH, JSON.stringify(all, null, 2));
}
const lastPostByIp = new Map(); // ip -> epoch ms (simple anti-flood)
function commentsBody(flash) {
const comments = loadComments().slice().reverse();
const list = comments
.map(
(c) => `
Questions, methodological objections, artifact reports, pointers to literature —
all welcome. The most valuable comment here is "your effect is explained by X". Comments are public.
Nothing on this page (or this site) is investment advice.
Leave a comment
${flash || ""}
${list || '
No comments yet — be the first.
'}
`;
}
app.get("/comments", (req, res) => {
let flash = "";
if (req.query.ok) flash = `
No comments yet — be the first.
'}