How we test — and the full results grid

This is the transparency page: the complete grid behind every guide on the site, plus exactly how each number is produced. Nothing is cherry-picked — the nulls and the near-misses are shown alongside the few robust results.

How each number is produced

Event study, look-ahead protected. For each release we measure the market's change from the prior close to the close at the end of the holding window. Nothing uses information that was not available before the release.
A baseline, always. The release-window move is compared against the unconditional distribution of same-length moves on all days — what a "normal" stretch looks like. Without a baseline, every release looks dramatic.
Two separate questions. - Direction — a one-sample t-test of the signed move against zero. Can you predict which way the market goes? - Size — the average absolute release-window move divided by the baseline (a ratio), tested by bootstrap. Does the market move more than usual, regardless of direction?
Multiple-testing correction. Run one grid of ~70 cells and a few will look "significant" by chance. We apply a Benjamini-Hochberg false-discovery-rate correction (q=0.05) across the grid. A result is only called robust if it survives that correction.

The full release-day grid

Each cell is the average absolute move on release day versus a normal day. Bold = survives the multiple-testing correction. Everything else — including plenty of numbers above 1.0 — does not, and we treat those as not robust. Yields are in basis points; the dollar and oil in percent returns; the ratio is unit-free.

Release	N	3M	2Y	5Y	10Y	30Y	2s10s	USD	WTI	Brent
CPI	37	×0.46	×1.54	×1.68	×1.58	×1.29	×1.09	×1.26	×0.96	×0.84
PPI	36	×0.52	×1.48	×1.47	×1.33	×1.08	×1.23	×1.01	×0.78	×0.85
NFP (jobs)	36	×0.62	×2.35	×2.28	×1.97	×1.48	×1.37	×1.38	×1.01	×1.11
PCE	35	×0.44	×0.80	×0.96	×0.98	×0.92	×0.81	×0.67	×0.62	×0.63
GDP (low N)	12	×0.27	×1.37	×1.33	×1.12	×0.91	×1.12	×0.54	×0.85	×0.79
Retail sales	38	×0.45	×1.29	×1.42	×1.32	×1.08	×0.97	×1.04	×0.88	×1.00
Jobless claims	150	×0.44	×1.06	×1.17	×1.15	×1.01	×1.12	×1.02	×0.77	×0.85
FOMC	24	×0.52	×1.61	×1.74	×1.45	×0.99	×1.08	×0.80	×0.86	×0.91

What survives, and what doesn't

Only three cells are robust: the jobs report at the 2-, 5- and 10-year. That is the entire robust set. CPI comes close — its 5- and 10-year gaps clear significance when tested on their own — but neither survives correction across the grid.
A lot looks elevated but isn't. CPI, PPI, retail sales and the FOMC decision all post ratios well above 1.0 — and none survive the correction. This is the whole point of the page: large-looking numbers are common; robust ones are rare.
The 3-month bill is calmer on release days, not louder (every ratio below 0.63) — it is anchored to the policy rate, not to a single data print.
Direction is unpredictable everywhere. Not one cell in the grid shows a robust signed direction. Across every release and every market, you cannot predict which way the move goes.

Why most of the grid is "not robust"

Of the ~70 cells, after correction the breakdown is: a handful robust, a middle band that is nominally significant but fails multiple-testing (we flag these for review, not publication), and the rest genuine nulls. We deliberately do not spin each cell into its own page — that would be low-value duplicate content. The grid lives here, on one page, in full.

Why the 2023–2025 window?

A fair question about this grid is why it covers 2023–2025 (every release date in the window, and nothing outside it) rather than two decades of history. The choice is deliberate, and it is a trade-off worth stating openly:

Baselines must be comparable. Our test compares release-day moves against "normal" days. The 2010s were a near-zero-rate world where the 10-year barely moved on any day; 2022–2025 is a very different volatility regime. Pooling them into one baseline would blend two different definitions of "normal" and could manufacture — or mask — release-day effects that aren't really there.
The cost is a shorter sample. Fewer events per release means less statistical power, and it means our direction findings in particular describe a single policy regime. That is exactly why we refuse to promote in-sample directional tilts (FOMC, CPI-dollar) to "rules," and why GDP's 12 events are flagged as low-N.
The road map is regime-split, not regime-blind. The right way to use older data is to test each era separately and compare — "does the CPI-day volatility bump exist in the 2000s, the 2010s, and the 2020s?" — rather than pooling everything into one average. That extension is planned; until then, every claim on this site should be read as a statement about the current regime, tested honestly within it.

Caveats

Adjacent maturities move together, so the effective number of independent tests is smaller than the cell count; the correction is conservative.
GDP has only 12 advance releases in the window (marked low N) — shown for completeness, not relied on.
No surprise-versus-forecast conditioning: this is the release-day reaction itself, not a reaction split by how far the number missed expectations. (Forecast/consensus data is proprietary and outside our public-domain sources.)
Historical statistics for informational purposes only, not financial advice. Results may vary with sample, period, and baseline definition.

Read the write-ups

Which U.S. data releases move the Treasury curve?
Do U.S. data releases move the dollar or oil?
Or any single-release guide from the home page.

Sources

All series are U.S. government public domain (17 U.S.C. §105), retrieved via FRED:

Treasury constant-maturity yields and the 10Y-2Y spread, Federal Reserve Board — DGS3MO, DGS2, DGS5, DGS10, DGS30, T10Y2Y.
Broad U.S. Dollar Index, Federal Reserve Board — DTWEXBGS.
WTI and Brent crude, U.S. Energy Information Administration — DCOILWTICO, DCOILBRENTEU.
Release dates: U.S. Bureau of Labor Statistics (CPI, PPI, jobs report), Bureau of Economic Analysis (PCE, GDP), Census Bureau (retail sales), Department of Labor (jobless claims), Federal Reserve (FOMC).