Can a frozen causal spatial correction improve ETAS across tectonically distinct regions without degrading CSEP calibration?
A 365-day, four-region shadow experiment. Forecasts are locked before each target day, then scored against earthquakes that occur afterward.
One test, four tectonic settings
The same causal architecture is evaluated without regional refitting in California RELM, New Zealand CSEP, the Chilean subduction corridor, and Japan C. Results are reported both by region and as a pre-specified pooled analysis; no region can be removed after activation.
This system compares statistical spatial rate forecasts. It does not predict individual earthquakes or provide public-safety guidance.
Four forecast lines
| Line | Role | What changes |
|---|---|---|
| ETAS | Frozen benchmark | Regional background plus triggered seismicity. |
| Safe incumbent | Conservative correction | Renewal-frailty and supported-neighbor residual structure. |
| Fixed spatial expert | Learned challenger | Moves probability mass only inside the highest-ETAS support. |
| Evidence-gated | Primary live challenger | Blends safe and fixed lines only after prospective evidence accumulates. |
Every line preserves the ETAS total expected count for that region-day. The experiment therefore isolates spatial allocation skill rather than winning by forecasting more earthquakes overall.
A causal, one-way evidence gate
Only completed prospective days can affect a later forecast. Fixed-versus-safe event log ratios accumulate in shadow mode, and the next forecast weight is:
The state begins at S=0; retrospective evidence is never imported. Until the BF20 hurdle is crossed, the gated line is identical to the safe line. BF20 is an operational hurdle, not an unconditional 5% false-activation guarantee.
What is measured
Primary: information gain per earthquake
Positive IGPE favors the challenger; negative IGPE favors ETAS. Because daily rate mass is preserved, the paired compensator difference is zero.
Calibration: CSEP tests
| Test | Question | Published view |
|---|---|---|
| N-test | Was the observed earthquake count consistent with the forecast count? | Two-sided Poisson p-value |
| L-test | Was the observed spatial likelihood typical under the model? | Lower-tail simulation p-value |
| R-test | Did the challenger outperform ETAS on paired likelihood? | Log-likelihood ratio and ETAS-null p-value |
Prospective integrity
- Backfill is prohibited; a missed forecast is never recreated after the target begins.
- The 365-day clock starts only after one atomic four-region publication.
- The seven-day interval is catalog settlement, not a pause in forecasting.
- Any model, feature, threshold, geometry, or gate change requires a new protocol identity.
Known limitations
The method was developed after retrospective outcomes in all four regions were inspected, so those results are not independent confirmation. Catalog completeness and agency continuity differ by region; Japan’s prospective targets use ComCat and are explicitly separated from the historical JMA development catalog. Sparse days can make daily CSEP values volatile, so the final interpretation must include event count, regional coverage, downtime, and uncertainty—not IGPE alone.