SoCal Earthquake Forecasting -- TFT
Temporal Fusion Transformer forecasting monthly earthquake activity (event count, max magnitude, mean depth, Gutenberg-Richter b-value) per 0.25-degree grid cell across Southern California. Trained on the SCEDC catalog. See the project README for full methodology: https://github.com (see repo README for the canonical link).
Version: initial (trained 2026-07-25) Weighted validation MAE: 0.8944
This is the seed model: trained natively (GPU workstation) before the AWS SageMaker retraining pipeline existed, then bootstrapped in as the initial champion. There is no prior champion to compare it against. Every subsequent retrain the pipeline runs is compared against this baseline (or whichever model has since become champion), and only promoted here if it's a strict improvement.
Validation MAE by target
| target | MAE |
|---|---|
| event_count | 0.7458 |
| max_mag | 0.3245 |
| mean_depth | 2.5796 |
| b_value | 0.0761 |