Guidance only, not an official warning. This is an experimental outlook made by one person. It is not issued by, and has no connection with, the UK Met Office or any official body. For official advice follow the Administration.

Method, and how far to trust it

An experimental outlook. It is guidance, not a warning, and the Administration's advice always comes first.

In short. The wind outlook has real skill for the big lee-side wind storms, but it has been tested on only six damaging events. The rain outlook is a weak guide: it barely beats chance, it missed the February 2024 flood, and it does not consider snowmelt on the high slopes. A "Normal" rain rating does not mean flooding cannot happen.

What is being forecast

Wind: the chance of a damaging downslope windstorm in any 48-hour window. This applies only to the lee side of the peak, which means the Settlement and Potato Patches side (west, clockwise through north, of the summit). It is not a forecast for the whole island, and it is not a forecast of gusts at a particular place or minute.
Rain: a relative risk of a heavy-rain flood or landslide event in the same 48 hours. It is not a percentage chance.

Global weather models cannot represent the island's mountain, so neither the extreme winds nor the heavy rain appear in them directly. The outlook instead looks for the large-scale set-ups that went with past events, using what the global model can see.

The wind index

For each 48-hour window the outlook takes two numbers from the forecast at a point 37.0°S, 12.5°W, just north-west of the island:

Low-level wind speed. The highest 850 hPa wind speed (about 1.5 km up) during the window, counted only while the wind blows from the east through south (90–180°), the directions that put the Settlement in the lee.
Shear. The highest amount by which that 850 hPa wind exceeds the 700 hPa wind (about 3 km up), in the same directions. A strong low-level wind under much lighter air aloft is the set-up in which mountain waves become trapped, grow large and break on the lee side. See the explanation and model plots.

The two numbers go into a logistic regression, a simple statistical formula that turns them into a chance between 0 and 100%. It was fitted to the 1980–2026 ERA5 reanalysis, using damaging wind events from 2006 onward (June 2008, May 2012, July 2019, November 2019, July 2026, and the mixed rain and wind event of June 2006). Before 2006 the island's news record is too patchy to say that nothing happened, so earlier days were not used as "no event" days. The May 2001 hurricane was kept aside as an independent check.

A backing-wind term (wind turning anticlockwise with height) was also tried. It added less than shear did, so it is not used.

How well it worked in testing

Each year was left out in turn, the model refitted on the other years, and then used to predict the left-out year. That gives an honest test.

Of the 6 damaging events, 5 reached at least 20% (4 of those 5 reached 50%). The one miss, May 2012, was an easterly gale that the reanalysis placed from the east-north-east, outside the 90–180° window. Widening the window to north-east scored it at 17%.
The independent May 2001 event scored 83%.
Across 20 years of days, a 5% alert would have produced about 1.2 false alarms a year, a 20% alert about 0.2 a year.
Statistical skill over always quoting the average chance: 0.28 (0 is no skill, 1 is perfect).

An "analogue" method (finding past days with similar numbers) did worse, so there is no sign the formula is missing something obvious. Six events is still very few. Treat the percentages as rough.

Where the forecasts come from

The same two numbers are calculated from the ECMWF AIFS deterministic forecast (a machine-learning weather model, 0.25° grid), which runs four times a day; the outlook uses the 00:00 and 12:00 UTC runs. Each run is used out to 72 hours. For each run the outlook gives the chance for the 48 hours starting now, 12 hours from now and 24 hours from now. The previous runs for the same windows are kept, which is why the page can show how the chance has been changing from run to run.

The confidence rating comes from how much the last two runs disagree about the same window: within 10 percentage points is high, within 25 is medium, more than that is low. Those limits come from the one forecast-era storm (July 2026), where successive 6-hourly runs typically moved by 3 points and, one time in ten, by 12 or more. They are provisional.

Calibration against the forecast

Training on a reanalysis and running on forecasts is not like for like, and this is the weakest link in the wind outlook.

Compared with the reanalysis over 19 months, the forecast winds are close (correlation about 0.94 to 0.96) but read a little low, by 0.3 to 0.7 m/s. In the strongest 10% of cases the error grows from about 4 m/s at the start of a run to about 6 m/s at 24 hours.
A correction for each lead time was tried. It did not improve things on held-out data, so it is not applied.
Only one damaging wind event, July 2026, has happened since the forecast archive begins (March 2025), so the forecasts cannot yet be calibrated against events. In that storm the forecast chance peaked at about 76%, compared with 84% from the reanalysis. It first passed 20% about 62 hours before the peak and 50% about 38 hours before. So treat the percentage as a minimum.
In about a thousand sampled 48-hour windows with no event since March 2025, the outlook raised no 5% alert. That is encouraging, but those were mostly quiet days and the sample is thin.

As the archive grows, each live forecast is kept so it can be checked against what happens, and verified reports from this site (see Reports) will help.

The wind levels

Level	Chance	Meaning in testing
Low	under 5%	Quiet days
Watch	5–20%	About one false alarm a year at 5%
Elevated	20–50%	About one false alarm every five years at 20%
High	50% and above	Every past damaging event reached at least 54%, except the May 2012 miss and the June 2006 mixed event (31%)

The 50% boundary and the wording are judgement calls, not tested thresholds.

The rain outlook, and why it is weak

What it does. It takes the largest 24-hour amount of rain that the forecast puts on that same sea point during the 48 hours. Over about 20 years of windows (2006 to 2026), 10% of windows had more than about 22 mm; in those, heavy-rain events were about 3 times as common as usual (1.8% of windows against a usual 0.6%). In the top 2%, above about 40 mm, they were about 5.5 times as common (3.3%). That is why the page shows "Elevated" and "High" as relative risks.

Even "High" means most such days pass without a flood. A 3.3% event rate still means about 97 in 100 such windows had no heavy-rain event.

How weak is it?

Over the eight heavy-rain events tested (June 2006, May 2006, January 2007, February 2011, October 2016, two in April 2017, February 2024), the best version scored a ranking skill of 0.63 (0.5 is chance). Its probability for none of the eight events reached 5%.
Set to catch events, it flagged only 2 of 8 at the lowest alert level, and gave about 11 false alarms a year doing it. It has no measurable skill over quoting the average chance.
It would have rated the February 2024 flood, the worst in 12 years, "Normal". The model showed only about 7 mm of rain that day, and the large-scale wind was weak.
The reanalysis put April 2017's "over 90 mm" day at about 27 mm, so the model rain is often much lower than what falls on the mountain.

What was tried and did not help. Moisture flux, vertically integrated moisture transport, how long a moist west-to-north flow persisted (to catch frontal waves stalling a warm conveyor belt over the island), and longer rain totals. Only the 24-hour rain amount carried any signal. Several events fell outside the west-to-north flow, including the July 2019 rain, which came in the south-easterly gale, so a flow-based rain rule drops some real events.

What it does not consider, and what matters for floods and landslides:

Snowmelt and rain on snow on the high ground at the top of the catchments. Snow has been reported on the slopes above about 500 m, and melt adding to heavy rain is not in the outlook at all.
How wet the ground and gulches already are, loose debris and blocked channels, and slope stability.
The island's own mountain: the rain is taken at a sea point just north-west of the island, so the extra rain the slopes squeeze out is not represented.
Narrow, short bands of heavy rain and thunderstorms, which a global model places badly. Forecasts of these bands are not very predictable.
Rain gauge and stream observations on the island. None are used yet.

Please read the rain level as a weak nudge to take care, never as reassurance.

Strengths and weaknesses at a glance

Wind outlook

Strengths

Based on the set-up we understand physically (strong low-level flow, trapped waves).
Caught every major lee-side storm in testing, 2001 included.
Few false alarms. Runs agree well, so it settles early.
Gives several days' notice in the one forecast-era case.

Weaknesses

Only six damaging events to learn from.
Forecast winds read low, so it understates.
Does not give timing, place or gust speed.
Misses storms from outside the east-to-south directions.
Cannot yet be calibrated on live events.

Rain outlook

Strengths

Honest: shown as a relative risk.
Real, if modest, signal in model rain over 24 hours.

Weaknesses

Barely better than chance.
Missed the February 2024 flood.
Many false alarms for each event caught.
No snowmelt, ground wetness, debris or local rain.
Model rain underestimates heavy falls on the mountain.

Data and sources

Training: ERA5 reanalysis (ECMWF, 1980 to 2026). Forecasts: ECMWF AIFS Single open data. Events: reports on tristandc.com and islanders' records. The island news archive is not complete, so small events are missing. Verified reports sent to this site will be added.

This page describes an experiment. It is not a warning service and it comes with no guarantee.