Testing the forecasting approach
Did higher-ranked regions separate from lower-ranked regions?
This page is the technical appendix. It tests whether the regional ordering separated large-event areas, followed the observed ranking, and stayed stable when inputs changed slightly. The rules were fixed before this test and were not adjusted to make the 2026 results look better.
Technical detail · 23 regions · eight labeled large events · one season
How the original and updated inputs compare.
Vintage / metricOrdinal v2Continuous v2.1
Pre-season AUROC0.9830.933
Pre-season Spearman0.7340.634
July 1 AUROC1.0000.967
July 1 Spearman0.7860.724
Top-five perturbation stability0.5440.974
The ordinal panel carries documented outcome non-blindness. The continuous inputs are computed by fixed rules from public data. These are validation results for one 23-region season—not a universal performance guarantee.
Québec · the coverage case
A pre-season call, preserved.
The March 16 QC-N record was RED at a continuous relative index of 0.775. By the July 21 evidence snapshot, CIFFC/SOPFEU corroboration recorded a merged Nord-du-Québec complex above 40,000 hectares within the analytical envelope.
Coverage is the product.
After late-June rain, the July 1 snapshot moved QC-N to ORANGE at 0.701. The rules were not retuned. Snapshot-versus-narrative disagreement remains visible.
Two more anchor stories
The promotion—and the refusal.
CA-NE. Continuous EDDI evidence promoted Lassen–Modoc from ORANGE to RED pre-season. The region later produced the roughly 16,188 ha Biscar fire. The record preserves the a-priori call without claiming post-hoc credit.
AFR-SAV. The archived window held 15,189 detections in 24 hours, while the qualifying incident check was not met. Satellite activity alone does not confirm a ≥10,000 ha wildfire incident.
Fidelity ledger
Eight flips. Every one adjacent.
VintageRegionOrdinalContinuousScores
preON-NWREDORANGE0.745 → 0.702
preSK-NREDORANGE0.745 → 0.666
preBC-NORANGERED0.682 → 0.721
preCA-NEORANGERED0.679 → 0.725
julQC-NREDORANGE0.848 → 0.701
julMB-NORANGERED0.828 → 0.759
julNL-LABORANGEYELLOW0.663 → 0.471
julCA-S4YELLOWORANGE0.654 → 0.554
Scope & limits
How to interpret this record.
- Absolute scores are uncalibrated relative indices.
- The 23-region validation covers one season with eight large-event labels.
- Seven inputs remain ordinal; two sampling points are sparse for heterogeneous macro-regions.
- Snow depth is a documented SWE proxy, not water-equivalent snowpack at native resolution.
- The southern-hemisphere July-forward phase is structurally early for its main fire season.
- The R12 thresholds were fixed a priori but have not yet received a threshold-sensitivity study.
- Global-forward scores already include R12 where applied; alignment is not independent validation.
- Analytical regions overlap, so regional evidence counts cannot be summed as independent totals.
- Display geometry is approximate and incident-source coverage is uneven across jurisdictions.