RECAST-Modern: Spatial Neural Point Process Earthquake Forecasting
A RoPE-Transformer + graph-attention (k-NN spatio-temporal) neural temporal point process, architecturally inspired by RECAST (Dascher-Cousineau et al., https://github.com/keliankaz/recast) but extended with a genuine spatial output head -- upstream RECAST is temporal-only. Forecasts weekly regional 7-day earthquake hazard lambda(t,x,y) for Southern California from the full pySTP microseismicity catalog (M>=0, 2011-present).
Research/experimental forecast -- not an official USGS or Caltech/SCSN product.
Promoted: 20260808T001613Z (trained 2026-08-08) Held-out NLL: -2.3846 (previous champion: -2.4547)
Held-out NLL by term (new vs. previous champion)
| term | new | previous champion |
|---|---|---|
| time | -3.0734 | -3.0622 |
| magnitude | -0.3160 | -0.3096 |
| spatial | 1.0048 | 0.9171 |
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support