Experiment V · Reporting
The Retrograde Audit
Mercury retrograde gets blamed for delayed flights, dropped calls and broken deployments. The windows are real and computable to the hour. Whether anything actually clusters inside them is a separate question, and an answerable one.
What retrograde actually is
Mercury does not reverse. It appears to because Earth, on a faster inner-track orbit, overtakes it — the same effect as a train sliding backwards when another passes it. The apparent reversal is viewing geometry, and it is as predictable as a timetable.
That makes the windows genuinely real. They are not a belief, they are a projection effect you can compute. Which is exactly why the question of whether anything clusters inside them can be settled by counting rather than by argument.
The first connected dataset
The US Bureau of Transportation Statistics reports every scheduled domestic flight operated by its reporting carriers. The event was fixed before analysis: ArrDel15 = 1, meaning arrival at least fifteen minutes behind schedule. Every other reported flight remains in the exposure denominator.
After adjusting for month and day of week, the 95% interval still crosses a rate ratio of 1. This dataset does not show a detectable difference in delayed-flight rates during Mercury retrograde.
| Analysis | Adjusted rate ratio | 95% confidence interval | Inside-window delay rate | Outside-window delay rate | Two-sided p |
|---|---|---|---|---|---|
| Astronomical window | 1.001 | 0.998–1.004 | 21.38% | 21.25% | 0.6787 |
| Window + 7-day shadow | 1.032 | 1.029–1.035 | 21.75% | 21.06% | < 0.0001 |
Poisson regression with log scheduled flights as the exposure offset; month and day of week are prespecified covariates. The seven-day shadow is the single prespecified secondary analysis. Pearson dispersion diagnostic: 368.42.
Overdispersion sensitivity check
The daily counts are far more variable than an ideal Poisson process. A non-confirmatory quasi-Poisson sensitivity check therefore scales the model-based uncertainty by the observed Pearson dispersion. It does not replace the filed model; it shows how strongly that model's narrow interval depends on its variance assumption.
Astronomical window: dispersion-scaled 95% interval 0.944–1.061, p=0.9828
Window + 7-day shadow: dispersion-scaled 95% interval 0.980–1.086, p=0.2287
What the source data look like
Monthly delay rates move substantially across seasons. Ember bands mark months containing at least one retrograde day; the visible variation is why comparing two raw averages without calendar controls would be inadequate.
These dates are computed, not copied
Retrograde dates are easy to find and impossible to check by looking at them. So they are derived here from JPL orbital elements — Mercury's and Earth's positions, converted to the longitude you would measure from the ground — and the derivation is tested against facts known independently of it: successive windows must be one synodic period apart (115.88 days), three to four must fall in a year, each must run about three weeks, and Mercury must never appear more than about 28° from the Sun. If the arithmetic were wrong, those checks would fail.
src/lib/retrograde.ts ↗ · the checks ↗
Windows
| Station · retrograde begins | Station · direct again | Days |
|---|---|---|
| 15 Mar 2025 | 7 Apr 2025 | 23 |
| 18 Jul 2025 | 11 Aug 2025 | 24 |
| 9 Nov 2025 | 29 Nov 2025 | 20 |
| 26 Feb 2026 | 20 Mar 2026 | 23 |
| 29 Jun 2026 | 23 Jul 2026 | 24 |
| 24 Oct 2026 | 13 Nov 2026 | 20 |
| 9 Feb 2027 | 3 Mar 2027 | 22 |
| 10 Jun 2027 | 4 Jul 2027 | 24 |
| 7 Oct 2027 | 28 Oct 2027 | 21 |
| 24 Jan 2028 | 14 Feb 2028 | 21 |
| 21 May 2028 | 14 Jun 2028 | 24 |
| 19 Sept 2028 | 11 Oct 2028 | 22 |
ALL DATES UTC · COMPUTED, NOT COPIED
About the shadow periods
Astrological practice often extends the window by one to two weeks either side. Stretching a window until an effect appears is the textbook garden of forking paths, so the primary analysis uses the astronomical window only, and exactly one secondary analysis uses the window plus seven days either side. Both get reported whatever they show.
Dataset, reduction, and provenance
The ingestion job downloads each official monthly ZIP, records its SHA-256, and reduces flight rows to daily counts. The compact snapshot is committed with the site; a scheduled workflow checks for newly released BTS months and redeploys only when those counts change.
BTS table profile ↗ · official download page ↗ · retrograde-audit.csv ↓ · archive manifest ↗ · scripts/ingest-retrograde-data.ts ↗
- Data coverage
- 2024-01-01 → 2026-06-30
- Source archives
- 30
- Snapshot generated
- 21 Aug 2026
The filed protocol
SHA-256 c958bbd02db65d3d60dda61871d1c0a5a57fbaa5c718a158c8a5f913b7799e1a
Known ways this could mislead
- Retrograde windows are not spread evenly across the calendar, and neither are flight delays. A retrograde falling in January would otherwise be credited with January weather, so month and day of week are covariates from the outset.
- Outage counts measure monitoring as much as they measure outages. Only sources whose methodology is stable across the whole period qualify.
- Testing several datasets is several chances at a spurious result, so Holm correction is applied across all of them — and every dataset examined is listed, analysed or not.
- A rate ratio can be statistically distinguishable from 1 and still far too small to matter. The rate ratio is the headline number here, not the p-value.
- Daily delays cluster during weather and network disruptions, so a Poisson model can understate uncertainty when the dispersion diagnostic is well above 1. The model-based interval remains the filed result, but it should not be read as a causal estimate.
Reuse this instrument
Share the current result or embed its compact, accessible summary. The card reads the same public aggregate as this page and carries the current N.
Embed this result
Continue through the research
These links connect the claim to the test, method, or neighboring history that can challenge it.