Classic AI models cry wolf at every minor tremor, inflating their accuracy numbers. Talivio works differently. Our system is deliberately under-confident — it never triggers an alert when the evidence is uncertain.
Say goodbye to false alarms. Know only the moments that truly require action.
| Feature / Metric | Traditional Institutions (USGS) | Classical Academic AI | Talivio (Regional ML) |
|---|---|---|---|
| Core Logic | Pure Statistics (only knows aftershocks) | Black-box Classification (memorises) | Physics-Informed AI (Spatial/Temporal) |
| Spatial Resolution | 10–50 km (very broad) | Province or region level | H3 Res 8/9 (0.7 km² pinpoint) |
| Time Horizon | Usually 24 hours | Single & uncertain | 7, 30 & 90-day dynamic probabilities |
| False Positive | Medium | Very high (up to 80% false alarms) | Near-zero (Precision@0.7 = 91%) |
| Probability Calibration (ECE) | None | Usually not measured | 0.032 (excellent calibration) |
| Brier Skill Score (BSS) | 1.5–3% (reference) | Negative (worse than random) | 23.9% (far above industry standard) |
Talivio's performance compared to published earthquake forecasting systems worldwide. Our AUC values use hard same-region negatives — the most rigorous evaluation protocol. Values auto-update as models are retrained.
| System | Metric | Value | ASS† | Window | Negatives | Validation |
|---|---|---|---|---|---|---|
| ETAS Italy (OEF-Italy) | Area Skill Score | 0.7 | 0.700 | 1-day | same_region | Prospective |
| RELM California (CSEP) | Probability Gain vs Random | 10 | — | 5-year | same_region | Prospective |
| CSEP California (2011–2020) | IGPE vs Benchmark | 0 | — | 5-year | same_region | Prospective |
| FCN Deep Learning (California) | Area Skill Score | 0.882 | 0.882 | 15–90 day | same_region | Retrospective |
| DeVries 2018 (Google Brain) | AUC | 0.849 | — | static | global_mixed | Retrospective |
| ETAS Japan (CSEP daily) | CSEP Pass Rate | 0.9 | — | 1-day | same_region | Prospective |
| EEPAS New Zealand | IGPE | 0.64 | — | 3-month | same_region | Prospective |
| Talivio v2 | AUC + ASS (Molchan) | 0.619–0.994 | 0.219 | 30-day | Hard same-region | Retrospective + Prospective |
† ASS = Area Skill Score (Molchan diagram, normalized [0,1]). FCN Deep Learning=0.882, ETAS Italy=0.70 from published papers.
Talivio ASS computed live from verified prospective forecasts — updates automatically as outcomes are confirmed.
Talivio ranks #3 among ASS-reporting systems.
* AUC world ranking based on honest evaluation: hard same-region negatives (model learns "will this fault rupture?" not "is this a fault?").
Most published ML AUC >0.90 use geographically distinct negatives.
8 regions, 127 backtest years. Full methodology →
Our system has been tested with Walk-forward Validation and a 30-day purge gap protocol to prevent data memorisation.
Her bölge için ayrı bir ML modeli eğitilir. Negatif örnekler aynı tektonik bölgeden seçilir — bu gerçekçi AUC değerleri sağlar (0.62–0.99 arası, coğrafi yanlılık yok).
Yeşil band = eğitilmiş model mevcut · AUC = expanding-window backtest sonucu (hard negative)
This system provides statistical probability forecasts. Earthquake time, location and magnitude cannot be precisely predicted scientifically. Probabilities are based on historical patterns and physical models; they do not imply certainty.