Cosmic-ray fluctuations reveal disturbed space weather, but fail a storm-warning test

Looking for a warning in cosmic rays

Geomagnetic storms occur when disturbances in the solar wind transfer energy into Earth’s magnetic environment. They can disrupt satellites, navigation and power systems. Galactic cosmic rays—high-energy particles travelling through the heliosphere, the region influenced by the Sun—also respond to interplanetary magnetic structures. Their intensity can fall during disturbances in events called Forbush decreases.

Haoyang Li, Zongyuan Ge and Zhaoming Wang investigated whether differences between cosmic-ray monitoring stations could provide advance warning of storms. Their revised arXiv preprint, posted on November 6, 2026, reports an average enhancement before storms, but no useful standalone warning performance from the simple alert method tested. The supplied record does not establish peer-reviewed publication.

Comparing stations against a past-only baseline

The researchers analysed hourly, pressure-corrected neutron-monitor counts from seven stations between 1995 and 2019. The stations span both hemispheres and respond differently to cosmic rays because of their locations and particle-energy sensitivities. Hours with fewer than four available stations were excluded.

A central methodological concern was avoiding a baseline that benefits from knowledge of the future. Cosmic-ray counts change over the solar cycle, so an unsuitable reference level could create misleading apparent precursors.

For each station, the team fitted a reference level using geomagnetically quiet days from the preceding year only. They also fitted daily and half-daily oscillations using the preceding 27 days. Each station’s count rate was then expressed as a percentage departure from its reference.

Two measures captured disagreement across the network. Inter-station dispersion, called A_disp, is the difference between the largest and smallest station departures. Mean-removed variance, A′_var, measures how widely those departures scatter around their network average. Neither reconstructs the directions from which cosmic rays arrive.

Removing the network average matters: the study found that an uncorrected measure based on squared departures was dominated by the common cosmic-ray decrease, rather than differences between stations.

A pattern emerges when storms are averaged

The team aligned observations on catalogued storm sudden commencements, or on main-phase onset when no such commencement was available. They then averaged the measurements within storm-intensity groups.

Both indices were elevated in the analysis window from 96 to 24 hours before the alignment time, with larger average amplitudes for stronger storms. For example, average dispersion increased from 1.70 percentage points for minor storms to 2.53 for extreme storms.

To check whether the pattern exceeded ordinary background variation, the researchers examined 96 isolated events. They compared these with randomly selected times from the same calendar years, excluding nearby storms. Both indices exceeded the quiet-control envelope across much of the pre-storm window.

That finding is an average association, not a countdown. The enhancement was broad and gradual; its appearance several days before the alignment time does not establish several days of reliable warning for an individual event.

The alert test exposes the difference

The practical evaluation asked whether an unusually large index value could trigger a useful warning. Thresholds were calculated from the preceding year, and alerts were permitted only when the current geomagnetic index indicated that a storm was not already underway. An alert counted as a hit if it occurred within 48 hours before main-phase onset. That interval was an evaluation choice, not demonstrated warning time.

Across 983 storm onsets, the representative 95th-percentile threshold detected 23.8% of storms using dispersion and 24.1% using mean-removed variance. The false-alarm ratios were 73.4% and 72.0%, respectively: roughly seven in ten alert episodes were false alarms.

A Molchan diagram compared missed storms with the fraction of time covered by warnings. The alert systems showed no positive skill against the random-alert benchmark. Lowering thresholds increased detection, but also increased warning coverage and false alarms.

Intensity estimates were limited too. Among 208 storms with catalogued sudden commencements and minimum Dst at or below −50 nanotesla, peak pre-commencement dispersion had a Spearman correlation of −0.28 with minimum Dst. More negative Dst means a stronger storm, but the intensity groups overlapped substantially.

Useful diagnostics, with important limits

Station coverage changed over time, and some stations had multi-year gaps. Restricting the analysis to all-seven-station hours retained the average intensity ordering, but weakened the dispersion result. The authors therefore regard mean-removed variance as the more robust measure.

The sample also mixes storms driven by coronal mass ejections with those associated with high-speed solar-wind streams, without separating their performance. Only 17 storms were extreme, limiting confidence for that category.

The authors suggest that these indices may track broad magnetic disturbances without directly constraining the southward interplanetary magnetic-field component important for storm severity. That explanation remains to be tested. Their conclusion is narrower than a forecasting breakthrough: cosmic-ray differences diagnose disturbed conditions, while reliable warnings would require complementary information such as solar-wind and magnetic-field measurements.