A seismicity-rate alarm from the public JMA catalog came before 6 of 61 Japanese eruption onsets in 1999-2024 with an uncertain gain, and about 99% of its alarms had no eruption

We tested if shallow earthquake rates in the public JMA unified hypocenter catalog increase before Japanese eruption onsets. We used JMA files for 1998-2024 and the Smithsonian GVP eruption list (snapshot 2026-11-09) for 89 Holocene volcanoes and 61 confirmed onsets with a known day at 16 volcanoes. The planned alarm was on when the previous 30 days had at least 5 events at 10 km depth or less within 10 km, and at least 3 times the count expected from the year before. The alarm was on before 6 of 61 onsets (9.8%, 95% volcano-bootstrap interval 2.9-21.1%). It was on for 3.9% of days at the same volcanoes, which gives a probability gain of 2.5 (0.8-5.1). Only 11 of 1,407 alarm episodes at the 89 volcanoes (0.8%) were followed by an onset within 30 days. Four of the six hits had alarms only 1-6 days before onset. A threshold-free rank score gave 0.59 (0.49-0.71), where 0.5 is no skill. Low-frequency events gave no skill (1 of 61 hits). Eight sensitivity rules gave point gains of 2.0-3.7, and three had intervals above 1 without correction for multiple tests. For 22 onsets, mostly at island volcanoes, the catalog had no nearby event before onset, so a precursor cannot be seen there. Main limitations: the unified catalog does not contain all volcano-network earthquakes, the sample is small, and onset dates can be uncertain. Code, plan, aggregate results, figures and input hashes are included. Raw JMA and GVP data and GVP-derived record tables are not included and are regenerated by scripts.

volcanic eruption forecasting · seismic unrest · Japan · JMA hypocenter catalog · Global Volcanism Program · false alarm · probability gain · low-frequency earthquakes · null result