How every number was made
Everything on this site comes from NOAA's public hurricane data files and short shell scripts in the repository, which anyone can rerun.
world-atlas and us-atlas packages.The files use the comma-separated ATCF "a-deck" (forecasts) and "b-deck" (best track) format. Field positions were read from the files themselves: cycle, model ID, forecast hour, latitude, longitude, maximum wind, pressure, storm type. Wind speeds are 1-minute sustained winds in knots. Rows repeat once per wind-radii threshold, so we keep the first row for each aid and forecast hour. A value of 0 kt means either no value or a dissipated system, and we drop it.
Fig. 1 shows the early (interpolated) aids NHC forecasters have at forecast time: OFCL HCCA IVCN HFAI HFBI HWFI HMNI CTCI AVNI AEMI UKX2 CEM2 SHIP DSHP LGEM GDMI NNIC OCD5. The UK and Canadian aids arrive as 6-h-old (UKXI, CEMI) or 12-h-old (UKX2, CEM2) versions depending on the cycle; this cycle has the 12-h-old ones. 12-h-old duplicates of aids that also have a 6-h-old version (HFA2, HWF2, GDM2) are left out to reduce clutter. Model families follow how each model works, as described on the model key.
The official forecast shown is OFCL for the 12Z cycle, issued with the 15Z advisory. Many published intensity plots show OFCI instead, the previous advisory's forecast interpolated forward. For Isaias at 12Z, OFCI and OFCL both peak at their starting 105 kt. OFCL and the early aids all start from 105 kt at 12Z.
To put an empirical uncertainty band on the official forecast (Fig. 1), we verified every NHC official forecast from 2023–2025 against the final best track. A forecast counts only when the system was a tropical or subtropical cyclone at both the start and the verifying time, roughly as NHC does in its own verification. Error is official minus observed. The table lists absolute-error quantiles.
| Lead time | Forecasts | Mean abs. error | Mean bias | Median | 67th pct | 90th pct |
|---|---|---|---|---|---|---|
| +12 h | 930 | 5.1 kt | +0.8 | 5 | 5 | 10 |
| +24 h | 829 | 7.6 kt | +0.9 | 5 | 10 | 15 |
| +36 h | 737 | 9.2 kt | +0.5 | 5 | 10 | 20 |
| +48 h | 654 | 10.6 kt | +0.2 | 10 | 10 | 25 |
| +60 h | 575 | 11.5 kt | −0.1 | 10 | 15 | 25 |
| +72 h | 502 | 12.5 kt | +0.4 | 10 | 15 | 30 |
| +96 h | 382 | 13.4 kt | +0.9 | 10 | 15 | 30 |
| +120 h | 293 | 14.4 kt | +0.4 | 10 | 20 | 30 |
The quantiles are coarse because NHC forecasts and best-track intensities are rounded to 5 kt. These are all-storm errors. Errors for storms approaching land, or undergoing rapid intensification, can be larger. Treat the band in Fig. 1 as "typical official error," not a probability forecast for Isaias. As a check, the mean absolute errors are close to the magnitudes NHC reports in its official verification.
Model spread is the sample standard deviation of the forecast intensity across a fixed set of eight aids, all present every season from 2023 to 2025: HFAI HFBI HWFI HMNI CTCI AVNI DSHP LGEM. A cycle and lead time count when at least six of the eight have a forecast. Fig. 5 compares Isaias's 12Z spread with every cycle that began at hurricane strength (official intensity at least 64 kt at the start): 371, 365 and 344 cycles at +24, +48 and +72 h.
Choices that would change the answer: a different aid set, including weaker storms (across all cycles, Isaias ranks in the 72nd and 26th percentiles at +24 and +48 h; fewer than six core aids reach +72 h this cycle), or conditioning on cycles near landfall. We did not condition on landfall. Spread in knots naturally shrinks as storms weaken over land.
The "What has already verified" table scores every forecast from earlier cycles whose valid time has passed, using the operational best track through 12Z 9 Oct (tools/early_verify.awk, output in data/early_verify.csv). Averages cover a homogeneous sample: lead times of +12 to +48 h, restricted to the cycle and lead time pairs where all 14 compared aids have a forecast (28 pairs). Track error is great-circle distance in nautical miles. The 48-hour positions and intensities quoted from the 12Z Wednesday cycle (and, for comparison, the four cycles before it, re-scored against the latest best track) use raw (non-interpolated) output: GDMN, HFSA, HFSB, HWRF, HMON, CTCX, AVNO, AEMN. The table rows are generated by tools/verify_table.sh. The sample is small and the best track is preliminary.
Requires bash, curl and GNU awk, for example Git Bash on Windows.
tools/build_data.sh al092026 ISAIAS 2026100912 tools/build_data.sh al092026 ISAIAS 2026100906 tools/spread_climatology.sh 2023 2024 2025 tools/official_errors.sh 2023 2024 2025 tools/summarize.sh al092026 2026100912 awk -F', *' -v bfile=data/raw/bal092026.dat -f tools/early_verify.awk data/raw/aal092026.dat > data/early_verify.csv