Methodology

1 · What HazardNet publishes

Each forecast unit is one published record: district, hazard class, severity, confidence score and 7- and 15-day horizons, stamped with the forecast date of the run that produced it.

• Outlooks for all 64 districts and eight hazard classes.
• A severity value in two tracks, a skill track and a physics track, each labelled honestly.
• Published alert levels where an alert policy applies.
• A validation scorecard and a live freshness record.

2 · How to read a severity value

Severity is a continuous value with two tracks reported side by side. The skill track is the system's own score for the class it selected; the physics track is an independent estimate. They can disagree, and when they do, that is shown rather than smoothed away. Confidence scores are uncalibrated.

3 · Horizons and lead time

Outlooks are published at 7 and 15 days. Both are planning windows, not warnings. The longer horizon is the more uncertain of the two.

4 · Freshness, staleness and failure

Every surface shows the forecast date behind it. A stale run is labelled as stale; a missing run is shown as missing: no number is interpolated to fill a gap.

5 · What is validated: and what is not

Forecast skill is not yet validated against observed outcomes. The published scorecard reports detection counts for five historical episodes and states what those numbers cannot support. No calibration accuracy is claimed.

6 · Methods are research-private

Model code, dataset collection procedures, training and benchmarking are research-private. Only results and outputs are published: on this site and in the public repository. For research collaboration or licensing, use the contact page.