Cycle-Scale Forecasting of Galactic Cosmic Ray Intensity: A Skill Assessment and The Role of Recurrent Solar Wind Structure
Year: 2026
Type: preprint
Abstract: Long-term forecasting of Galactic Cosmic Ray (GCR) intensity is conventionally assessed by how closely a model reproduces the observed record rather than by whether it outperforms a naive alternative. Using monthly Oulu and Moscow neutron monitor data for 1976 to 2025, together with heliospheric current sheet (HCS) tilt, Interplanetary Magnetic Field (IMF) magnitude, solar wind speed and the aa geomagnetic index, we evaluate cycle-scale Cosmic Ray Intensity (CRI) forecasting against a persistence benchmark under strict rolling-origin validation. Here all smoothed predictors constructed causally so that no information from after the forecast issue time enters any forecast. A conventional regression on CRI level, evaluated at matched forecast horizons, attains out-of-sample correlations as high as 0.91 yet fails to beat persistence at every horizon. It acquires a systematic bias of approximately − 4% when extrapolated from the calibration period into a later heliospheric state characterised by weaker IMF. Reformulating the target as the change in CRI over the forecast horizon largely removes this bias. The point-estimate skill is negative at six months, where persistence is difficult to beat, and turns positive from twelve months onward, reaching 0.36 at twenty-four months. All 95% bootstrap intervals nevertheless include zero and Diebold-Mariano tests do not reject equal predictive accuracy with persistence. A 27-day recurrence index derived from the aa index, hitherto known as a descriptive correlate of cosmic ray flux, contributes negligible aggregate skill at twelve months despite being individually significant there, and raises the aggregate point-estimate skill at eighteen and twenty-four months, with a standardised coefficient rising monotonically with lead time from 0.09 to 1.20. The orthogonal component of the index, after removal of the variance shared with the mean-field predictors, retains most of that contribution. No statistically robust incremental forecast skill from aggregate transient injection was detected, on either the Richardson and Cane near-Earth Interplanetary Coronal Mass Ejection (ICME) list or the CDAW SOHO/LASCO catalogue, under an acceptance criterion fixed in advance. Calibrating the neutron monitor network against AMS-02 daily proton fluxes accepts seven of twenty-two station and detector combinations and yields median response rigidities from 10 to 72 GV, with Thule, Oulu and Moscow degenerate near 10 GV despite cutoffs differing by a factor of eight, and with atmospheric depth rather than cutoff governing the response. Forecast skill is highest and most variable among monitors near 10 GV, falls to a lower level near 19 GV, and falls essentially to zero at 72 GV, consistent with useful cycle-scale skill being concentrated below approximately 20 GV. Nested cross-validation confirms that adaptive choice of the predictor specification yields no improvement over the fixed model, while retaining the recurrence index in every training fold at the longer horizons. No skill score attains significance at 95%. For the present validation design the estimated detectability threshold is about 0.4, and the observed 0.36 falls just below it, so the results do not provide sufficient evidence to reject the persistence benchmark. We conclude that the results provide suggestive evidence that recurrent solar wind organisation contains information beyond the mean interplanetary field at horizons of eighteen to twenty-four months, that no comparable evidence is found for aggregate transient injection, and that the available record is too short to establish differences of this magnitude at conventional confidence.
Source: Research Square
Authors: V. K. Mishra, Praveen Tyagi
Language: English
Subfield: Astronomy and Astrophysics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225406990