Research › Plain-language summary
Roberts (2026) — Pre-Earthquake Geomagnetic Quietude and Event-Day IMF Variability: Replication Evidence (2010 to 2016)
In the 10 days before major earthquakes, the geomagnetic environment was measurably quieter than matched non-earthquake days. That pattern held across 350 earthquakes and 26,250 comparison windows, was repeated 12 different ways with different control groups, and was not driven by any single year or by differences in solar cycle phase.
The finding is counterintuitive. Most people would expect an earthquake, if it had any relationship to space weather at all, to be preceded by a spike or storm in geomagnetic activity. The pattern found here runs the other direction: sustained quiet, not elevated activity, is what appears more often in the days leading up to a major earthquake.
Earthquake data came from the NOAA National Centers for Environmental Information (NCEI) Significant Earthquake Database: 350 events of magnitude M≥5.0 across 328 unique calendar dates, covering 2010 through 2016. Space weather data came from the NASA OMNI-2 database, augmented with data from the SILSO solar monitoring center and the University of Oulu neutron monitor.
85 different space weather measurements were tracked simultaneously — everything from solar wind speed and magnetic field components to geomagnetic activity indices recorded at ground stations worldwide. These were not hand-picked; they were every measurement that had enough valid data in the baseline year.
The primary pattern tested, called the broad quiet signal, required at least 30 of those 85 measurements to be simultaneously suppressed across the 10 days before an earthquake. A metric counted as suppressed if it was below its year-specific average on at least 8 of the 10 pre-event days, and its 10-day average also fell below the year mean. This definition requires broad, sustained suppression — not one metric on one day, but a general quieting across many independent measurements over the full pre-event window.
26,250 control windows were drawn from non-earthquake dates across 2010 to 2025, matched to the same time of year as each earthquake to control for seasonal and solar cycle variation.
In plain terms: the broad quiet pattern appeared about 29% more often in the 10 days before a major earthquake than before matched non-earthquake days. The probability this is a random fluctuation is 0.3%.
Two additional measurements confirmed the same direction independently:
These are not derived from the same pattern-counting method — they are raw averages of independent measurements that confirm the same directional result by a completely different approach.
A single statistical result means very little on its own. The value of this study is that the main finding was subjected to 12 independent tests — different control groups, different year ranges, different sample compositions — and held in every one.
Earthquake windows compared against matched control days drawn from 2010 through 2025.
Controls restricted to 2010 to 2019, avoiding any comparison across different solar cycle phases.
Controls restricted to the exact same seven-year span as the earthquake sample.
The strictest test: control windows from the same calendar year as each earthquake, confirmed to contain no earthquake dates. Eliminates solar cycle phase as an explanation entirely.
All 2010 earthquakes dropped; test re-run on the remaining 297 events.
All 2011 earthquakes dropped, including the Tōhoku M9.1 event.
All 2012 earthquakes dropped. Effect size strengthened slightly with this year removed.
All 2013 earthquakes dropped.
All 2014 earthquakes dropped. The weakest of the seven leave-one-out configurations, but still significant.
All 2015 earthquakes dropped.
All 2016 earthquakes dropped. This produced the strongest leave-one-out result of all seven configurations. See the 2016 section below for why this year is complicated.
Some calendar dates had two earthquakes. When duplicates were removed and only the highest-magnitude event per date was kept (328 unique dates), the finding held.
21 different patterns were tested in total. The paper reports what did not show up, which matters as much as what did.
Single-day quiet signals on the day before or the day of the earthquake did not replicate. Single-day spikes in activity on earthquake day did not replicate. Quiet-to-spike transition patterns on earthquake day did not replicate. General counts of notable readings on the day before or day of did not replicate. The signal is specific to sustained breadth of suppression across the full 10 days — not anything dramatic on earthquake day itself.
The paper is upfront about one anomaly. When 2016 earthquake windows were compared against cross-year controls, the quiet pattern appeared less often before the earthquakes (22.00%) than before the controls (29.25%). That is the wrong direction.
The cause was identified: 2016 was a geomagnetically elevated year. The cross-year control pool drew mostly from quieter years, so control windows had a higher quiet rate not because they were associated with earthquakes but simply because those other years were calmer overall. Of the 15 control years used for 2016 comparisons, 13 had higher quiet rates than the 2016 earthquake windows.
The paper does not hide this. It presents 2016 as an unresolved boundary condition and identifies it as the reason same-year controls matter.
Exploratory — not yet independently replicated
Among the 115 earthquake windows that already showed the 10-day quiet pre-signal, a separate pattern was examined for earthquake day itself. The north-south component of the interplanetary magnetic field (Bz) had been fluctuating very little during the quiet pre-window — then on the day of the earthquake, the rate of Bz polarity reversals jumped sharply.
This result is robust within the analysis but it is exploratory. It has not been independently replicated in a second, non-overlapping dataset. The paper reports it as a candidate pattern requiring further testing — not as an established finding.
The study makes no claim that geomagnetic conditions cause earthquakes. The mechanism, if any exists, is entirely unknown and is explicitly stated as not established by this data.
The analysis covers seven years (2010 to 2016) and earthquakes of magnitude M≥5. Whether the finding holds for larger events, other time periods, or specific geographic regions has not been tested.
Under full Benjamini-Hochberg correction across all 21 tested patterns in the main control design, the primary result carries q = .063, just above the conventional .05 threshold. Under the strictest control design (same-year earthquake-screened), it drops to q = .011. The paper presents this transparently.
Earthquake data from the NOAA National Centers for Environmental Information (NCEI) Significant Earthquake Database. Space weather data from the NASA OMNI-2 interplanetary database, augmented with International Sunspot Number data from SILSO (Royal Observatory of Belgium) and cosmic ray data from the University of Oulu neutron monitor. Statistical method: matched-set permutation test with Benjamini-Hochberg false discovery rate correction, 3,000 iterations.