Every night we check our forecasts against what the rivers actually did, using ten measures with targets we set ourselves. The numbers below are not hidden when they are bad.
Forecasts met 0 of 10 targets over the 7 days measured so far.
Be careful with these numbers. There have been only 26 times in this period when a river reached a flood mark, so a handful of events can swing the results. Forecasts we got right on a quiet day tell us little; the targets about flood marks matter most.
When a river reached a flood mark, had we already said it could, early enough to act?
17 of 26 river crossings of a flood mark were warned about in time. Warnings came a median 8.1 hours ahead.
Target: at least 80% warned in time: 6 hours ahead for the alert level and 12 hours for minor and major flood levels (2 hours on fast-rising rivers) · 26 cases
When we said a river could reach a flood mark, how often did it stay below it?
80% of "possible" warnings and 25% of "likely" warnings did not happen.
Target: no more than 60% of "possible" warnings and 25% of "likely" warnings fail to happen · 73 cases
When we show a chance of 3 in 10, the river should reach that mark about 3 times in 10.
5 of 7 groups of chances matched what happened.
Target: within 10 points of what really happened, for every chance with at least 10 cases · 4605 cases
How much closer our level forecast is than just saying the river will stay where it is now.
Closer than "no change" by 33% at 1 hour, 28% at 3 hours, 20% at 6 hours, 16% at 12 hours and 4% at 24 hours. Weakest kind of movement: 0%.
Target: 30% closer at 1 to 6 hours, 20% at 12 hours, 10% at 24 hours, and never worse than "no change" in any kind of river movement · 9392 cases
Whether the 12 hour forecast tends to be above or below what the river really did.
At 12 hours the forecast was 0.45 ft too high on average. The worst kind of movement was off by 1.46 ft.
Target: off by no more than 0.25 ft on average, in every kind of river movement · 1449 cases
Our forecast shows a range that should hold the real level 4 times in 5.
The real level was inside the range 79% of the time: 81% at 1 hour, 80% at 6 hours, 79% at 12 hours and 72% at 24 hours.
Target: between 75% and 85% of the time at every look-ahead · 9392 cases
When the river rose 10 cm or more in an hour, did the "Next hour" card say Rising?
73% of 177 fast rises were called Rising.
Target: at least 95% of fast rises called Rising · 177 cases
When the card said Rising, did the river go up?
70% of 534 Rising calls were followed by a rise.
Target: at least 75% of Rising calls followed by a rise · 534 cases
Whether the forecast ran every hour of the day, on enough gauges, and how often a faulty sensor reading got in the way.
Forecasts in 13 of 24 hours a day, from 40 gauges an hour. 1% of forecasts started from a reading that looked like a sensor fault and were left out of these checks.
Target: forecasts in all 24 hours of every day
Of the people who answered "Is this right?" on our pages, how many said the forecast matches what they see.
70% of 89 answers said the forecast matches what they see.
Target: at least 70% say it matches · 89 cases
On the 3 days with heavy rain, forecasts met 1 of 10 targets.
We improve the forecast one change at a time. Each change is tested beside the live forecast first, and only then goes live. Here is how the forecast scored while each version was live.
From 29 Oct 2026
One-reading spikes and frozen sensors are cleaned out of the readings the forecast learns from, so a faulty gauge no longer distorts it.
No forecasts were checked while this version was live.
From 26 Oct 2026 · 2,525 forecasts
In storms the forecast reads the rain the satellite actually measured, timing rain beside the gauge and rain upstream separately.
From 26 Oct 2026 · 158 forecasts
Heavy rain over the catchment now lifts the 12 and 24 hour forecasts, and the stated range widens when a storm is under way.
From 24 Oct 2026 · 166 forecasts
The next 1, 3 and 6 hours now come from the river's own recent movement (fitted on 11 months of readings) instead of extending the trend.
From 22 Oct 2026 · 667 forecasts
The first forecasts: a rain-based model for the next hours, and a simple trend line for gauges without a model.
Updated 29 Oct 2026. Checked from 22 Oct 2026 to 28 Oct 2026. Levels are in feet. The numbers behind this page