Hurricane · next 24 hours

Rapid intensification: the chance in the next 24 hours

Rapid intensification means a tropical cyclone’s maximum sustained winds rise by 30 knots or more within 24 hours — the kind of change official forecasts find hardest to anticipate. For each active storm we publish the best available probability: NOAA’s own guidance where it exists, our model elsewhere. Each storm names its source.

Updated
Active storms
1
Sources in use
HazardPulse v8.3

Active storms

This forecast was issued 31 hours ago, so its 24-hour window has passed. A new one is published every 6 hours, after each forecast cycle; until then, treat these numbers as out of date.

Quality checks: Published with warnings. The forecast is published; a non-blocking check flagged the following. Full check record

  • Hashes missing for: model file, input data, receipt
  • Calibration on live forecasts not yet measured (too few closed forecasts with events)

For tracks, warnings and evacuation decisions, follow the National Hurricane Center, Central Pacific Hurricane Center and Joint Typhoon Warning Center. HazardPulse publishes research forecasts, not warnings.

Map of active tropical cyclonesEach ringed marker is an active tropical cyclone, labelled with its name and coloured by its chance of rapid intensification. The cards below give each storm.Xavier: 0.5%Xavier
Chance of rapid intensification, next 24 hours
  • under 1%
  • 1–5%
  • 5–15%
  • 15–30%
  • 30–50%
  • 50% or more

Xavier

Southern Hemisphere · SH042027

0.5%chance of rapid intensification, next 24 hours · source: HazardPulse v8.3
Strength
Tropical Storm: 55 kt maximum sustained winds, 989 hPa
Position
11.3°S 164.0°W, 1,700 km WNW of Papeete, French Polynesia
Forecast cycle
Where the number comes from
HazardPulse v8.3: NOAA publishes no rapid-intensification guidance for this basin

Where each number comes from

Atlantic, East and Central Pacific. NOAA’s own probability, read from the National Hurricane Center’s SHIPS text for the same forecast cycle: DTOPS as issued, or SHIPS-RII when DTOPS is missing. We chose it in a comparison written down in advance and run on 2020–2024, then scored it once on the 2025 season: it ranked a cycle that went on to intensify rapidly above one that did not 92.8% of the time over 601 forecast cycles and 33 storms, and beat NOAA SHIPS-RII, NOAA RIOC consensus and our previous model (v8.2) on the same cycles.

West Pacific, Indian Ocean and Southern Hemisphere, where no public rapid-intensification guidance exists, and any NHC cycle without usable SHIPS text: our v8.3 model. It was tested on 2,967 forecast cycles from 2025–2026 in the basins where it publishes (South Indian 1,302, West Pacific 1,078, South Pacific 440, North Indian 147), none of them used to fit it: it ranked a cycle that went on to intensify rapidly above one that did not 85.2% of the time. It is v8.2’s method refitted without the duplicate copies of storms that cross basins, which v8.2’s training rows held (amendment 15); on the same cycles its log loss was 0.1659 against v8.2’s 0.1656 (difference +0.0003, 95% interval −0.0007 to +0.0014), inside the margin of +0.002 written down before the test. That test read each storm’s inputs from the best track. In the West Pacific, Indian Ocean and southern hemisphere a storm’s recent track comes from the working best track (UCAR RAL’s real-time copy), the kind of track the model was trained on before its post-season revision. Scored from a single JTWC warning instead — which leaves 9 of its 17 inputs unknown, as it did until 2026-10-05 and still does if that file cannot be read — the 2025–2026 West Pacific cycles had a log loss of 0.227, against 0.240 for always forecasting the average (0.177 with every input). When a cycle’s best-track fix is not yet published, the warning is the analysis and the storm’s central pressure and its changes are filled with typical values: scored that way, 0.183. On West Pacific and South Indian storms, our satellite model J1 is shown beside it for comparison while it is tested on new forecasts against v8.2, the model it is built on; it would replace v8.3 only by meeting that test’s rule, written down in advance (its results and rule).

Read the full test

Our track and intensity forecast

For every storm the National Hurricane Center follows, HazardPulse issues its own track and intensity forecast, TC1, out to five days. It combines the computer models NHC publishes in real time. Each model is weighted by how it has actually performed: its errors are re-measured every six hours against where the storm turned out to be, so a model that is doing well this season counts for more, and a new one earns its weight as it proves itself. Model hurricane_tc1-589cd13ad9b6.

Track. Tested in advance on the 2023–2025 seasons, our track forecasts were off by 90.7 nautical miles on average over one to five days. NHC’s official forecast was off by 92.1, and its best automatic guidance (HCCA) by 97.6. Against the official forecast that is a tie at the confidence the test required (98.75%): the difference, −2.8 n mi [−8.2, 2.0], could be chance. Against HCCA, in a comparison the test reported but did not make a claim of (95% interval), it is ahead at every lead from 72 hours: −6.2 n mi [−11.1, −1.4].

Intensity. Our intensity forecasts were off by 12.8 knots on average, NHC’s official ones by 12.6: a tie (0.2 kt [−0.6, 1.1]). Intensity is where every forecast struggles most.

2026 so far. Against NHC’s official forecast on this season’s storms, scored against NHC’s operational best tracks: track −0.9 n mi [−3.9, 2.2], intensity 1.0 kt [−0.6, 2.2] (95% intervals). A season still in progress, so no claim is made from it.

TC1+O, the same weighting with NHC’s official forecast as one more member, is recorded too: its track was off by 90.5 n mi on 2023–2025, −3.1 n mi [−8.2, 1.4] against the official forecast, a tie at the confidence the test required (98.75%). It is not shown on the storm cards.

With our rapid-intensification model (TC2b, in test). Intensity forecasts miss most when a storm intensifies rapidly, which is exactly what our RI model forecasts. TC2b holds TC1’s winds over the next day to the middle of the range our RI model expects, and leaves TC1 alone wherever the two agree. On the cycles that went on to intensify rapidly, the 24-hour wind error fell from 25.1 to 20.0 kt on 29 cycles of 2026, the season it was tested on (NHC’s official forecast: 19.0 kt), and from 23.6 to 20.7 kt on 154 cycles of 2023–2025. Over every cycle the change is small: −0.15 kt at 24 hours [−0.62, 0.24] and −0.07 kt [−0.22, 0.05] over one to five days on 2026 (95% intervals), because most cycles are left as they were.

What the storm cards show: TC2b+O. The same rule applied to TC1+O, our forecast with NHC’s official forecast as one more member, which had the lowest intensity error of our forecasts. It improved on TC1+O in both seasons tested. On 2023–2025 its average intensity error over one to five days, 12.54 kt, was below NHC’s official 12.63 kt (difference −0.07 kt [−0.61, 0.47], which could be chance); on 2026 NHC’s official forecast leads, 11.05 against 11.27 kt (0.39 kt [−0.77, 1.20]). On the cycles that intensified rapidly, its 24-hour error was 20.5 kt against TC1+O’s 23.2 on 2023–2025 and 19.4 against 24.0 on 2026 (NHC: 18.5 and 18.3). It is shown beside our track forecast while it is tested on new forecasts; no claim against NHC’s official forecast is made.

Follow the National Hurricane Center for decisions. Ours is shown beside theirs so the two can be compared as the season goes on. The full test

Our experimental model

HazardPulse v10.1 is our own rapid-intensification model, built from NOAA’s intensity guidance, its rapid-intensification probabilities and the official forecast. Over the 2022–2025 seasons, each forecast only from earlier seasons, it scored better than DTOPS: log loss 0.144 against 0.165 (lower is better).

That is not yet proof. It is shown beside NOAA’s number, never instead of it, while a test written down in advance scores it on new forecasts; the verdicts are due on 1 Dec 2026 and 1 Dec 2027. Forecast cycles scored so far: 150.

The full comparison, including the variants under test