Experimental research
Earthquake research lab
Experimental researchOpen tests of the astronomical and temporal variables most often proposed as earthquake precursors. Each hypothesis is stated before the test is run, the method is fixed in code, and the result is published whichever way it comes out.
EXPERIMENTAL RESEARCH — NOT A VERIFIED EARTHQUAKE PREDICTION OR OFFICIAL WARNING.
Nothing on this page is a forecast, a warning, or an input to one. These tests ask a narrow question — whether the timing of earthquakes is statistically associated with a proposed variable — and the expected answer for most of them is no. Even a clear association would not be a prediction method: earthquakes cannot currently be predicted with useful precision in time, place and size, and no result here changes that.
This run
Computed 2026-10-27 13:40:00 UTC (2 h ago) · model schuster-gk74-bh-v1
Why the aftershocks are thrown away. Earthquakes are not independent. One magnitude 7 produces hundreds of aftershocks over days to years, all within a small area. Run a statistical test on the raw catalogue and you are counting a single physical episode hundreds of times, which manufactures significance out of nothing. Removing them costs statistical power and is the only way to get an honest answer.
Why the p-values are adjusted. Testing 4 hypotheses at the 5% level means roughly a one-in-five chance that at least one looks significant purely by luck. Every p-value below carries a Benjamini-Hochberg correction across the whole battery.
Source: USGS ComCat via Prakruti's reconciled catalogue · Retrieved: 2026-10-27 13:40:00 UTC · results are computed by a scheduled job, not per page view, so they are reproducible
Time of year and earthquake timing
Hypothesis (stated before the test was run)
Independent mainshocks are uniformly distributed across the calendar year. The alternative is a seasonal signal.
Why this is worth testing
Seasonal hydrological loading — snowpack, monsoon rainfall, reservoir level — produces surface stress changes that have been linked to seismicity in a few specific regions. Globally, any such effect should average out.
Method
As above, with the phase angle taken from the day of year.
Across 37,332 independent mainshocks, the phase angles depart from uniform (Schuster p = 0.0031, p = 0.0063 after correction), with a mean direction near 25° (0° = 1 January, 180° = early July). The effect is small: the resultant vector is 1.24% of its maximum, meaning the overwhelming majority of the timing is unexplained by this variable. A departure from uniformity is not a mechanism, is not causation, and cannot be used to anticipate any individual earthquake.
Observed distribution against the uniform expectation
Limitations of this result
A global test averages over both hemispheres, whose seasons are opposed, so a genuine regional seasonal effect would tend to cancel. A null result here does not rule out regional seasonality.
Reproducibility: model schuster-gk74-bh-v1; catalogue M4.5+ from 2011-01-01T00:00:00.000Z; Gardner-Knopoff declustering; Schuster test; Benjamini-Hochberg across 4 hypotheses
Lunar distance and earthquake timing
Hypothesis (stated before the test was run)
Independent mainshocks are uniformly distributed across the anomalistic (perigee-to-apogee) cycle. The alternative is clustering near perigee, when the lunar tide-raising force is strongest.
Why this is worth testing
The tide-raising force varies as the inverse cube of distance, so it is about 40% stronger at perigee than at apogee. If tidal stress modulates earthquake timing at all, the anomalistic cycle is where the effect should be largest.
Method
Identical to the lunar phase experiment, but the phase angle is derived from the Moon's distance mapped onto the perigee-apogee cycle, with 0° at perigee.
Across 37,332 independent mainshocks, the phase angles are consistent with a uniform distribution (Schuster p = 0.5068, p = 0.5068 after correcting for the other tests in this run). The resultant vector is 0.43% of its maximum possible length, which is what random scatter looks like. No association was detected.
Observed distribution against the uniform expectation
Limitations of this result
The anomalistic and synodic cycles beat against one another, so the two experiments are not independent and their p-values should not be read as separate pieces of evidence. Same catalogue and declustering caveats apply.
Reproducibility: model schuster-gk74-bh-v1; catalogue M4.5+ from 2011-01-01T00:00:00.000Z; Gardner-Knopoff declustering; Schuster test; Benjamini-Hochberg across 4 hypotheses
Lunar phase and earthquake timing
Hypothesis (stated before the test was run)
Independent mainshocks of magnitude 4.5 and above are uniformly distributed across the lunar synodic cycle. The alternative is that they cluster near some phase, such as new or full moon.
Why this is worth testing
The Sun and Moon raise measurable tides in the solid Earth, producing stress changes of a few kilopascals. Those stresses are thousands of times smaller than the stress drop of a typical earthquake, but they are real and periodic, so a small modulation of earthquake timing is physically conceivable and has been reported in some specific settings. This is the most plausible of the astronomical hypotheses, which is why it is tested first.
Method
Events are filtered to the magnitude of completeness, declustered with Gardner-Knopoff windows to remove aftershock dependence, and each remaining mainshock is assigned the lunar phase angle at its origin time. Schuster's test is then applied to those phase angles. The reported p-value is corrected for the number of experiments run.
Across 37,332 independent mainshocks, the phase angles are consistent with a uniform distribution (Schuster p = 0.1252, p = 0.1669 after correcting for the other tests in this run). The resultant vector is 0.75% of its maximum possible length, which is what random scatter looks like. No association was detected.
Observed distribution against the uniform expectation
Limitations of this result
The catalogue is global, so a regional effect confined to one tectonic setting could be diluted to invisibility. Declustering is conservative and may remove genuinely independent events. A statistically detectable modulation would still be far too weak to time an earthquake, and this test cannot establish causation.
Reproducibility: model schuster-gk74-bh-v1; catalogue M4.5+ from 2011-01-01T00:00:00.000Z; Gardner-Knopoff declustering; Schuster test; Benjamini-Hochberg across 4 hypotheses
Time of day and earthquake timing
Hypothesis (stated before the test was run)
Independent mainshocks are uniformly distributed across the solar day. The alternative is clustering at some hour of the day.
Why this is worth testing
Tectonic earthquakes have no known mechanism tied to local solar time, so this is included as a control. It also functions as a detector for catalogue artefacts: seismic networks sometimes report more events at night, when cultural noise is lower and small events are easier to detect.
Method
As above, with the phase angle taken from UTC time of day.
Across 37,332 independent mainshocks, the phase angles depart from uniform (Schuster p = < 1e-10, p = < 1e-10 after correction), with a mean direction near 253° (0° = 00:00 UTC, 180° = 12:00 UTC). The effect is small: the resultant vector is 3.24% of its maximum, meaning the overwhelming majority of the timing is unexplained by this variable. A departure from uniformity is not a mechanism, is not causation, and cannot be used to anticipate any individual earthquake.
Observed distribution against the uniform expectation
Limitations of this result
A positive result here would most likely indicate a detection artefact in the catalogue rather than anything geophysical, which is exactly why it is worth running alongside the astronomical tests.
Reproducibility: model schuster-gk74-bh-v1; catalogue M4.5+ from 2011-01-01T00:00:00.000Z; Gardner-Knopoff declustering; Schuster test; Benjamini-Hochberg across 4 hypotheses
What this run actually shows
Neither lunar hypothesis showed any detectable departure from uniformity in this run. Meanwhile the time-of-day control — which has no plausible geophysical mechanism — did show one, with an effect size of 3.24%.
That ordering is the finding. A control with no mechanism behind it produced a stronger signal than the hypotheses that do have one. The likeliest explanation is not that solar time moves faults, but that the catalogue itself carries structure: seismic networks detect small earthquakes more reliably at night, when human background noise falls, and the world's seismicity is unevenly spread across time zones. This is a measurement artefact wearing the costume of a discovery, and catching it is exactly what a control is for.
It is also why every p-value here is printed beside its effect size and never on its own. With tens of thousands of events, an effect far too small to matter will still clear any significance threshold you care to name.
How to read a null result
“No association detected” is not the same as “no association exists”. A test can only detect an effect large enough for the sample to reveal. With 37,332 independent mainshocks, a very small modulation would be invisible here — and would still be useless for anticipating an earthquake even if it were real.
Equally, “association detected” would not mean the Moon or a planet causes earthquakes. Association is not causation, and a detectable statistical effect can be orders of magnitude too weak to act on.
The minimum sample the engine will report on is 200 independent mainshocks above the magnitude of completeness (M4.5). Below that it refuses to publish a number, because a small sample can produce an impressive p-value that means nothing.
For the forecasting that does work — rates built on observed clustering and aftershock statistics, rather than on astronomy — see the forecast page.