Every earthquake this month, and the odds of the next one.

02 / is it any good

Tested on 780 earthquakes it never saw.

I fitted the model on every magnitude 6+ earthquake from 2000 to 2019, then froze it and asked it about every one from 2020 to August 2026. When it says 40%, an aftershock came about 40% of the time. Overall its forecasts were 27% more accurate than just using how often aftershocks have happened before.

03 / how it works

Aftershocks fade in a pattern. The first day says how strongly.

1. 01

   The record

   Every magnitude 4.5+ earthquake worldwide since 2000 from the USGS catalogue: 187,392 of them.

2. 02

   The sequences

   Each magnitude 6+ that is not itself the aftershock of a bigger one, with every magnitude 5+ within a distance that grows with its size.

3. 03

   The pattern

   Aftershocks thin out with time in a well-known way (Reasenberg and Jones, 1989), and bigger earthquakes have more. I fitted that pattern's four numbers to 2000 to 2019.

4. 04

   The first day

   Sequences differ a lot, so each one's own strength starts unknown and is updated by its first day's count, trusted about a third of the way. That weight was chosen on the training years too.

What comes next: the same four numbers fitted region by region, since a subduction zone and a mid-ocean ridge do not behave alike, and forecasts that update hour by hour through the week rather than once at the end of day one. The live forecasts above use exactly the tested rule: magnitude 6+, aftershocks of 5+, days 1 to 8.

Read the case study →