Research › Plain-language summary
Roberts (2026) — Pre-Referendum Solar Wind Density Variability and Geomagnetic Deviation Signals: A Cross-Window Replication Study of Referendum Events, 2010 to 2026
Two space weather signals showed a repeatable association with referendum dates in a global event inventory spanning 2010 to 2026. The same signals held specifically for referendums while failing under identical conditions for treaties and coups tested in parallel. Effect sizes are small — neither signal is large enough to predict individual events. What makes the finding worth reporting is structural persistence across a large number of independent filters.
This page explains what was found, how it was tested, and what it does and does not mean.
The study tested two space weather signals against a corrected inventory of global referendum events. The inventory was drawn from Wikidata and cleaned before any analysis began: scheduled elections were excluded from the source pull, and placeholder dates (dates where Wikidata had assigned January 1 as a default when no specific date was known) were removed. The final cleaned inventory contained 8,407 event rows across all political event types.
Within that, the referendum subtype contained 313 events in the 2010 to 2017 window and 387 events in the 2018 to 2026 window.
The two space weather signals:
How much the density of charged particles streaming from the sun was fluctuating on a given day. Specifically, the daily maximum of a six-hour rolling variance of solar wind proton density. A higher value means the density was swinging more than usual that day.
How far daily geomagnetic activity (measured by the Kp index) deviated from its long-term historical mean, expressed in standard deviation units. A higher value means that day's geomagnetic activity was further from normal than usual, in either direction.
Both signals were elevated before referendum dates compared to matched control days drawn from the same time period. The analysis was split into two completely independent time windows and the test was run separately in each.
Solar wind density variability at T-14:
An AUC of 0.5 would mean no difference from chance. Values above 0.5 mean referendum dates tended to follow days of elevated solar wind density variability. Both windows cleared correction for multiple testing (the q values reported above are false discovery rate corrected).
Geomagnetic deviation at T-9:
The second signal cleared strict false discovery rate correction in Window 1 and held in the correct direction in Window 2. Window 2 cleared the nominal significance threshold but not the stricter corrected threshold. The signal is treated as secondary for this reason. It later reappeared as a clean survivor in a broader validation analysis described below.
This is the most structurally important part of the result. The same two signals were tested simultaneously on three political event subtypes drawn from the same inventory: referendums, treaties, and coups.
Neither signal replicated across both windows in treaty events (175 events in Window 1, 67 in Window 2) or coup events (36 in Window 1, 43 in Window 2). The signals held only in referendums.
This is significant because treaties and coups are both political event types whose scheduling is driven by factors that have no obvious connection to space weather. That neither signal appeared in either type — under the exact same test applied to the same time periods — means the finding is not simply a feature of all political events. It is specific to referendums, and that specificity is the primary structural constraint on the result.
The primary signal (solar wind density variability at T-14) survived a large number of independent credibility checks. Each one addresses a different way the result could be an artifact rather than a real pattern.
The signal was first identified in a broader framework covering thousands of global events across multiple categories. Referendums were a subset tested afterward with the lead held fixed from the earlier work.
Regular elections happen on predictable calendar cycles, which could create false timing patterns unrelated to space weather. They were removed from the inventory before any testing.
Wikidata sometimes assigns January 1 as a default date when the actual date is unknown. Those rows were identified and removed before testing.
The analysis was restricted to referendum events within the broader politics category, with treaties and coups tested in parallel as controls. The signal held in referendums and failed in both other subtypes.
The full 2010 to 2026 span was split into two non-overlapping periods — 2010 to 2017 and 2018 to 2026 — and the test was run separately in each with no information shared between them.
The 14-day pre-event lead was fixed in advance from earlier work and was not adjusted to fit either window. It appeared at the same lead in both.
Higher solar wind density variability was associated with upcoming referendum dates in both the earlier and later time windows. The direction did not flip.
When many statistical tests are run, some will appear significant by chance. The Benjamini-Hochberg correction accounts for this. The signal cleared that correction in Window 1 (q = .024) and Window 2 (q < .001).
Referendum dates were compared against non-event days drawn from the same time frames, not against a theoretical average. The result means referendum dates had higher solar wind density variability than matched non-event days.
Some metrics can look significant simply because they carry information from earlier in time that bleeds into the pre-event window. The metric was independently verified as clean in the pipeline's feature registry.
A later, separate analysis of the broader all-politics population applied stricter checks including parity, support, and autocorrelation correction. The same metric reappeared as a clean survivor in that analysis on a different event population.
The broader validation required the signal to appear at the same lead day in both windows. It passed that gate (confirmed at T-1 in both windows in the all-politics population).
A looser check asking whether the signal held anywhere within the same general lead range. Passing both the strict same-lead gate and the regime-band gate means the result was not fragile to exactly which day was tested.
The broader validation's proof file explicitly confirmed the metric as clean and non-latency, consistent with the earlier registry classification under stricter stage requirements.
The AUC values (.549 and .567) fall below a prespecified practical effect floor of AUC .580. The paper states this explicitly rather than treating a small effect as a large one. Small but persistent is the honest description.
The finding does not mean space weather causes referendums, or that referendums influence the solar wind. The causal direction is constrained in one important way: space weather does not respond to institutional referendum scheduling. Whatever the relationship is, the pre-event elevation in solar wind conditions cannot be attributed to referendums affecting the sun. The mechanism, if any exists, is entirely unknown.
The effect is too small to predict individual referendum dates. It cannot tell you when a referendum will be called. Independent replication in a second, unrelated event source is the appropriate next step before drawing stronger conclusions. The Wikidata event labels used here have not been validated against independent electoral records.
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. Event inventory from Wikidata. Statistical method: Mann-Whitney AUC with Benjamini-Hochberg false discovery rate correction.