Fire-weather forecast check: Southern California, Jan 1–8 2025

Tier 1 local demo. The browser streams a local sharded Zarr v3 store and computes every score live. Nothing is precomputed: matching and statistics are MET v12.2.2's own code, run in your browser.

Loading…

How to read this page

About 3 minutes. It walks through the page and can change the settings for you.

The question: which forecast model best predicted the fire weather around the January 2025 Los Angeles fires? Three models (AIFS, GFS, HRRR), 8 forecast days (Jan 1–8, 00Z), leads 0–48 h.
The controls below change every score on the page, live. For a like-for-like comparison use Truth: URMA and Verify on: URMA 2.5 km grid (common).
Fire-weather event: humidity at or below the RH slider and 10 m wind at or above the Wind slider, at the same place and time.
Reading scores: for errors (RMSE) lower is better; for bias, closer to 0 is better; for CSI, POD and FSS, higher is better (1 is perfect).
Sections: score by lead time (how error grows with forecast length), the 24 h event table, the ranking across all 8 days, and maps of where each model misses.

Truth Cycles Metric RH ≤ Wind ≥ Verify on Matching (MET) Neighbourhood Exactimation

Scored with MET v12.2.2 in WebAssembly: –. Neighbourhood width per grid: .

Fire-weather event = RH at or below the threshold and 10 m wind speed at or above it, at the same cell and time.

Score by lead time

Choose the score with Metric. Hover the chart for each model's value at a lead. Forecasts usually get worse as the lead grows.

AIFS (0.25°)GFS (0.25°)HRRR (3 km)

Fire-weather contingency table at 24 h

Counts of grid cells (or station reports). Hits: event forecast and observed. False alarms: forecast, not observed. Misses: observed, not forecast. Correct no: neither. POD is the share of observed events caught; FAR the share of forecasts that were false; CSI combines both.

Model	Hits	False alarms	Misses	Correct no	POD	FAR	CSI

Ranking across the case set

Each case (one cycle at one lead) is scored on its own and kept as the lines grid_stat or point_stat writes (SL1L2, VL1L2, CTC, NBRCNT). MET's stat_analysis aggregation then pools those lines over all 8 cycles. Uses the truth, matching, thresholds and neighbourhood set above; the Cycles selector does not apply. “Best in k/n cycles” counts the cycles a model wins when each cycle is aggregated on its own; ties count for every tied model.

Leads

Read the ranking with care. Eight cycles around one event are a small sample. ASOS truth has only basic QC. URMA uses HRRR as its first guess, so it is not fully independent of HRRR, and HRRR analysis is HRRR's own starting point. Matching and statistics are MET's own code; RH for AIFS and URMA is a ratio of MET's saturation vapour pressures (the ratio is ours), and the URMA decode's equality with MET's own GRIB2 decode is still open.

Where each model misses

Per-cell scores over the case set, with MET series_analysis. Uses the truth and matching set above (gridded truth only).

Statistic Leads

Provenance

Licences and source

MET (Apache-2.0, © UCAR/NCAR, DTC) is compiled to WebAssembly in metwasm/met.wasm. That file also statically links 26 routines from GSL 2.7.1 (GPL-3.0-or-later), which provide MET's normal confidence intervals, so met.wasm is distributed under the GNU GPL version 3.
Corresponding Source for met.wasm: MET-WASM build sources (shim, build scripts, JavaScript API), MET v12.2.2 at commit 58a0612c, GSL 2.7.1, and the licence texts: GPL-3.0, MET LICENSE. tools/met-wasm/BUILD.md inside the build sources rebuilds it byte for byte.
Data: forecasts and ASOS reports from dynamical.org (CC BY 4.0); URMA analyses from NOAA/NCEP (public domain), via the NOAA Open Data Dissemination program on AWS.