Archived snapshot from the 12Z Thu 8 Oct 2026 model cycle. Forecasts have changed since. See the latest analysis.
NHC model cycle 12Z Thu 8 Oct 2026 · central Gulf of Mexico
Physics-based, statistical and AI models are all forecasting the same Gulf hurricane. This page shows where they agree, where they split, and whether the split is unusual. It uses NOAA's public model file and three seasons of verification.
Nearly every model agrees on the shape. Isaias strengthens for about a day, then weakens fast between +36 and +48 h as NHC's track takes the center inland over the north-central Gulf coast. The models disagree on two numbers: how strong the storm is at the coast, and how fast it falls apart over land.
Isaias sat nearly stationary in the Bay of Campeche for about two days, first as a weak disturbance and then as a depression. It then turned east-northeast and intensified rapidly, from 35 kt at 06Z Wednesday to 70 kt at 06Z Thursday. That is 35 kt in 24 hours. Early on, the models disagreed about both the turn and the intensification. Some moved the system east quickly and strengthened it. Others kept it slow, weak and farther south. Those forecasts can now be checked against NHC's best track.
The 12Z Tuesday runs, made 48 h before Thursday's 12Z position, show the split clearly. DeepMind's raw run placed the center 6 nmi from where it was. HAFS-A and HAFS-B were within 11 and 24 nmi. Over every forecast so far that all the aids below share, the interpolated DeepMind aid has the smallest average track error, 16 nmi. NHC's official forecast is close behind at 18 nmi.
The same DeepMind run forecast 69 kt for 12Z Thursday, two days ahead. The observed intensity was 70 kt. HAFS-A forecast 65 kt. Across the shared forecasts, the smallest intensity errors belong to the corrected consensus (HCCA), the neural-network consensus (NNIC), DeepMind, NHC and the Navy's COAMPS-TC, all between 4 and 6 kt.
Nearly every aid was too slow to intensify Isaias. The global models ran the lowest, as expected at their resolution, and the statistical LGEM was 12 kt low. Two days out, the GFS, its ensemble mean and the track consensus kept the center too far south or too slow, by 130–190 nmi. COAMPS-TC, the best interpolated physics model on intensity, had the largest track error of the regional models.
| Aid | Type | Intensity error | Intensity bias | Track error |
|---|---|---|---|---|
| HCCA | Consensus | 4.1 kt | −2.8 | 27 nmi |
| NNIC | Machine learning | 4.4 kt | −3.4 | — |
| GDMI | AI (DeepMind) | 5.0 kt | −3.8 | 16 nmi |
| OFCL | NHC official | 5.4 kt | −4.6 | 18 nmi |
| CTCI | Regional physics | 5.7 kt | +4.2 | 79 nmi |
| HWFI | Regional physics | 7.2 kt | −7.2 | 36 nmi |
| HMNI | Regional physics | 7.8 kt | −7.6 | 53 nmi |
| SHIP / DSHP | Statistical | 7.8 kt | −6.6 | — |
| HFAI | Regional physics | 8.2 kt | −7.6 | 44 nmi |
| HFBI | Regional physics | 9.8 kt | −7.6 | 43 nmi |
| AVNI | Global physics | 11.2 kt | −11.2 | 40 nmi |
| LGEM | Statistical | 13.2 kt | −12.2 | — |
| AEMI | Global physics | 19.0 kt | −19.0 | 56 nmi |
Mean absolute errors at +12 to +48 h, over the 12 forecast cycle and lead time pairs that every listed aid shares (a homogeneous sample), scored against NHC's operational best track through 12Z Thursday. Statistical and ML aids borrow another model's track, so their track errors are not listed. This covers two to three days of one storm against a preliminary best track. It shows which models handled this storm's genesis and rapid intensification, not which models are better overall.
At +24 h the regional physics models span 87–101 kt, with the Navy's COAMPS-TC highest. The global physics, statistical and interpolated DeepMind aids sit between 75 and 84 kt. Regional hurricane models follow the storm with nested grids (about 2 km for HAFS) that can resolve the eyewall. Global models, with grid spacings of about 10 km or more, cannot, and they usually run weak at the peak. NHC's 95 kt is near the top of the range.
At +48 h the DeepMind aid (GDMI) is at 23 kt, the regional physics models are at 29–43 kt and the GFS is at 39 kt. The UK global model and the Canadian ensemble mean are still at 62 and 66 kt, and the deterministic Canadian run (CMC2, not plotted) is also at 62 kt. SHIPS, a statistical model with no land in it by design, holds 67 kt. Its land-aware version, Decay-SHIPS, drops to 48 kt. The 19 kt between those two is the land effect as that statistical scheme models it, which makes the pair a useful control.
Some of the intensity spread is really track spread. A model that brings the center ashore earlier, or farther west, also weakens it earlier. At +36 h HAFS-A puts the center at 29.3°N 88.0°W and the DeepMind aid at 29.7°N 87.2°W, about 90 km apart (Fig. 4).
The DeepMind line on most spaghetti plots is GDMI, an interpolated aid. The raw model output (GDMN) arrives too late for the forecast cycle it would serve. NHC therefore takes the previous run and adjusts it to the storm's observed intensity at the start time.
For Isaias, the raw 06Z DeepMind run forecast 82 kt for 12Z. The observed intensity was 70 kt. After the adjustment, the 12Z aid runs 7–16 kt below the raw run through +48 h. The raw run peaks at 95 kt, the same as NHC's official forecast.
"AI is less aggressive" is partly an artifact of this adjustment. Compare raw and interpolated output before drawing conclusions about any model class.
Intensity forecasts over land depend on when and where the center crosses the coast. The tracks below cluster tightly toward the Alabama and western Florida Panhandle coast. Their small east–west and timing differences explain part of the +36 to +48 h intensity spread.
Not among the core models. The split comes from the models outside that core.
We took eight aids that ran in every season from 2023 to 2025: HAFS-A, HAFS-B, HWRF, HMON, COAMPS-TC, GFS, Decay-SHIPS and LGEM. For every Atlantic forecast cycle that started at hurricane strength, we computed their spread, the standard deviation across models. By that measure, Isaias's disagreement is ordinary at +24, +48 and +72 h.
Isaias stands out once the other models come back in. Adding the DeepMind aid, the UK global model, the Canadian ensemble mean and SHIPS raises the +48 h spread from 6.8 kt to 14.3 kt. That figure uses a different set of models, so it can't be compared to the histograms. It does show where the disagreement comes from: models that treat land very differently.
NOAA's global ensemble (GEFS) runs 31 versions of the GFS, each with slightly perturbed starting conditions and model physics. At this resolution its intensities are much too weak. At 06Z the members started between 38 and 72 kt against an observed 70 kt, so read the plume for timing, not strength. 25 of the 31 members lose 15 kt or more between +36 and +48 h on the 12Z clock. That is independent evidence that the timing of the move inland is fairly well constrained.
The European (ECMWF) model and ensemble are not in NOAA's public model file, so they are not shown. Google's model is itself an ensemble, but the public file carries one forecast per cycle.
This page is a snapshot of the 12Z Thursday cycle, committed to a public repository before landfall, and the commit history timestamps it. After the storm, each aid will be scored against NHC's best track from +12 to +72 h, on the questions this storm raises:
One storm can't settle which model is better. NHC's annual verification reports are where models get judged. This page is a worked example of reading the guidance.
This page is updated as new NHC model cycles arrive. Each earlier version is preserved unchanged.