The celestial studies

A pre-registered test of roughly thirty celestial and calendar hypotheses against equity returns, every one of them run through the same statistical harness, corrected for multiple comparisons across the whole family, and reported including the null results, which is most of them.

Where everything is, right now

Apparent geocentric positions from JPL DE440s, the same ephemeris the study computes its variables from. Scrub the slider to move through time and watch the aspect geometry form and break; bodies are drawn hollow while retrograde.

Nothing on this wheel is a claim. An aspect is two bodies at a particular angular separation, which is a measurable fact about the solar system. Whether that fact has anything to do with equity returns is precisely what the study below is built to find out, and the honest expectation is that it does not.

What happened

The harness ran over 14,013 sessions of the NASDAQ Composite, 1971-02-08 to 2026-09-04, 55.6 years. Every hypothesis got the same treatment: mean return by bucket against its complement, Newey–West HAC standard errors, a stationary bootstrap interval, and correction across the entire family at once.

Tests run

168

Hypotheses

26

p < 0.05, uncorrected

26

Survive BH, q < 0.10

5

Survive Bonferroni

3

Twenty-six results look significant if you never correct. Five survive Benjamini–Hochberg and three survive Bonferroni. That gap, between 26 and 3, is the entire lesson of this page, and it is why the leaderboard below has a line drawn across it.

By tier

Tier 3, the astrology, produced fewer apparent hits than chance alone would give you at this many tests. Tier 4, the calendar effects, produced the most, which is the result that says the harness can tell the difference, and the reason the control group is in the study at all.

The strongest result is real. The next two are not.

The Monday effect tops the table at −14 basis points a day, and it is a documented, replicated feature of equity returns going back decades. Finding it is evidence the machinery works.

Directly beneath it sit the Metonic cycle and the Saros cycle, both surviving correction, both apparently worth about 120 basis points over 21 days. Both are artefacts, and the study can show you why rather than asking you to take it on faith:

• The Metonic cycle is 19 years and the Saros is 18.03. In a 55.6-year sample that is 2.92 and 3.08 complete cycles. Splitting either into terciles does not sample a cycle, it cuts history into three long eras and asks whether they differed. They did.
• The two are supposed to be different hypotheses, yet their terciles label 89.5% of sessions identically, because 19 years and 18.03 years partition a 55-year span almost the same way. Two independent tests agreeing to three decimal places is not corroboration; it is the same test counted twice.
• Fourteen thousand daily rows make the standard errors look small. The effective number of independent observations of an 19-year cycle is three. No HAC lag and no bootstrap block length repairs that, and this is exactly the failure the Bangladesh line is about.

Complete cycles in sample, for every long-period hypothesis: metonic cycle 2.92, saros cycle 3.08, lunar nodes 2.99, saturn return 1.88, jupiter saturn conjunction 2.8, sunspot number 5.05.

The leaderboard

Every test, ranked by raw p-value, the one place on this site where anything is sorted by significance, and only to show what that sorting would have told you. Effects and intervals in basis points per day. MDE is the smallest effect this test could have detected at 80% power.

Not run

bradley_siderograph

Bradley's published weights (1948) could not be sourced. The pre-registration forbids inventing them and requires any reconstruction be labelled as one, so this hypothesis is reported as not run rather than run against guessed coefficients.

The hypothesis list

Ordered by prior plausibility, not by p-value. That ordering is the intellectual content of the study, and results will be displayed in it too, sorting a leaderboard by significance is how you turn a multiple-comparisons exercise back into a fishing expedition.

How every hypothesis will be treated

Identically. No per-hypothesis tuning, which is the whole design. Daily index log returns plus forward 5-day and 21-day windows; mean return by bucket against its complement; Newey–West HAC standard errors; stationary bootstrap confidence intervals with a block length chosen to preserve autocorrelation; and Benjamini–Hochberg alongside Bonferroni across the entire family. Both corrections get reported, because they disagree and the disagreement is informative.

Each hypothesis will also carry a power analysis: the minimum effect detectable at 80% power given the number of events that exist. For most of this list the honest finding will be that you would need more data than the market has produced, and that is the result worth publishing.

The Bangladesh line

When the leaderboard exists it will carry a horizontal rule at the corrected significance threshold, labelled the Bangladesh line, with one footnote beneath it:

In 1994, David Leinweber and Dave Krider searched a UN database of non-financial country statistics and found that butter production in Bangladesh explained 75% of S&P 500 variation from 1981–1993. Adding US cheese production took it to 95%. Adding the sheep populations of Bangladesh and the US took it to 99%. It was published as Stupid Data Miner Tricks. Everything above this line is butter.