| 23 Apr 2026
Critical Slowing Down in the Geomagnetic Ap Index as an Early Warning Signal for Major Geomagnetic Storms
Abstract. Geomagnetic storms pose critical risks to technological infrastructure, yet advance forecasting beyond 24 hours remains limited. Critical slowing down (CSD) — the tendency of complex systems to exhibit rising variance and autocorrelation as they approach bifurcation — provides a model-free early warning framework applicable to any system undergoing a critical transition. We apply CSD analysis to the daily planetary geomagnetic Ap index over 38 years (1987–2024; n = 13,880 days), computing a composite instability metric (rolling variance × |rolling first-order autocorrelation [AR(1)]|) across 30-, 60-, and 90-day windows. Receiver operating characteristic (ROC) analysis using DeLong standard errors demonstrates that the 30-day CSD index predicts major storms (Ap ≥ 30) with area under the curve (AUC) = 0.724 (95 % confidence interval [CI]: 0.705–0.744; Z-statistic = 22.49, p < 0.001) at a 30-day lead, significantly outperforming lagged Ap alone (ΔAUC = +0.132, p = 0.003). At the 90th-percentile CSD threshold, precision lift is 2.89× the base rate with Youden's J = 0.339. Pre-storm CSD elevation (−28.7 % above baseline) is in the expected direction but does not reach individual significance (Cohen's d = 0.22), while during-storm elevation is large (d = 1.36, p < 0.001). These results provide the first systematic demonstration that magnetospheric CSD dynamics, recoverable from a single surface index, contain actionable predictive information weeks before storm onset.
The paper introduces the concept of the Critical Slowing Down Composite Metric, as the product of the variance with the first autocorrelation, in windows of 30 days, claiming that it can provide predictive value for storm forecasting, as captured by the daily Ap index values. The idea is promising and relies on interesting notions from the theory of complex dynamical systems, while also being easy to implement and thus could have potential for forecasting applications, but I believe that the manuscript requires further clarifications in some points and possibly more thorough comparisons to better illustrate its predictive power.
Some points that require further clarifications:
Section 1: At the end of the Introduction: The author mentions the use of "lagged Ap" as a "natural baseline predictor". What exactly is meant by that? Is this the Ap lagged by one day? How is this used in the forecast?
Section 2.1: It is mentioned that "Daily solar indices (F10.7 flux, international sunspot number) were obtained from the same archives for covariate analysis". Are the results of this shown anywhere in the paper, or were used for any of the conclusions? Please elaborate (if necessary).
Section 2.2: Where other formulas tried than a simple product? Also, the absolute value in the AR is troubling, since it could indicate that strong negative correlations would have the same effect as strong positive ones. I guess that since only a time lag of one day was used, it's most likely that this never happens and the AR is either close to 1, during events, or close to zero (perhaps slightly negative, but still small in absolute value) during "noisy" intervals where no particular activity takes place. Thus, I assume that this does not create any problems, but I would like to see some more details by the author on the choice of the formula.
Section 2.3: The analysis looks thorough, but many of these metrics are unfamiliar to me. It would be overly optimistic to expect the reader to study five additional papers to understand the metrics used. I would recommend to either expand this section significantly by including the details of these measures or use fewer of them, but provide a short description and formulas. Also, a sentence or two are required to explain how the method works; how does it predict a future event? Is the detection based on a simple threshold and if its exceeded the method predicts an event for the next day? Or for some other time in the future? Or is it producing some probability of event occurrence? Can the author choose an optimal threshold value and give the confusion (or contigency) matrix of true/false hits/misses? As it's written, I'm afraid it will only make sense to statisticians.
Caption of Figure 1: The caption reads "Gray shading marks major storm epochs (Ap ≥ 30)". Earlier in the paper it is mentioned that there were 873 storm days, yet here I only see about 40 gray marks! Could it be that these are for the extreme events (Ap>100 instead of Ap>30)?
Figure 2: The legend in the top left panel (A) shows different values for the AUC that those mentioned in the paper and those shown in the table at the bottom right (F) of the same figure. Also, it looks like the lagged Ap is better than the other three methods! What is wrong here?
Sections 4 & 5: I find it rather difficult to believe that a method based only on Ap "predicts major geomagnetic storms at 30-day lead", without any information about the driver, i.e. the solar wind. Regardless of whether a storm is CME or CIR-driven the underlying cause of it is always the solar activity. Of course, the state of the magnetosphere is a major factor in how it will respond to the external driving, but predicting future events without any such information and based only on the disturbance level of the past month seems extra-ordinary!