Derivative-Aligned Anticipation of Forbush Decreases from Entropy and Fractal Markers

TL;DR

This paper presents a feature-based framework using information theory and fractal analysis to anticipate Forbush decreases in neutron monitor data, achieving high detection rates with significant lead times across multiple stations.

Contribution

The study introduces a novel, reproducible pipeline that leverages entropy and fractal markers for early detection of Forbush decreases without cross-station homogenization.

Findings

High detection rates with median leads of several hours.

Markers show sustained pre-event excursions, enabling early warning.

Method remains stable under parameter sensitivity analyses.

Abstract

We develop a feature-based framework to anticipate Forbush decreases in one-minute neutron-monitor records by tracking sliding-window invariants from information theory, scaling, and geometry. For each station we compute marker time series, including Shannon, spectral, approximate and sample entropy; Lempel-Ziv complexity; correlation dimension; and Higuchi and Katz fractal dimensions. Markers are smoothed with an exponentially weighted moving average and analyzed through within-station standardized first differences. Timing is referenced to an operational alignment time defined as the minimum of the smoothed count first difference, and marker leads are reported in minutes (negative values indicate anticipation). Station-level detectability is evaluated on a pre-alignment window using a robust z-score detector with bilateral threshold and persistence, without cross-correlation or hypothesis testing. We apply the pipeline to two FD episodes with broad station coverage (2023-04-23 and 2024-05-10; 28 stations each). Across events, a compact CORE panel shows consistently high detection rates and predominantly anticipatory lead distributions, with typical median leads on the order of several hours depending on the invariant and event. Lead dispersion across stations is substantial, with interquartile ranges commonly spanning a few hours, highlighting the need for station-wise criteria and distributional summaries rather than single-station inference. Representative marker trajectories confirm that early flagging corresponds to sustained pre-alignment excursions in marker differences, not tabulation artifacts. The approach is reproducible from open code, operates on native station units without cross-station homogenization, and remains qualitatively stable under sensitivity sweeps of windowing, smoothing, and detector parameters.