Methodology

Model development & learning

Transparent prediction-model evolution based on real Bampton flood data. We publish every accuracy figure and every algorithm change so the methodology can be audited.

Since 8 October 2026

How the prediction works now

On 8 October the front page switched to a new early-warning model. The range and the four colours look the same; what decides them has changed. The audio and the numbered sections below explain the two-model setup from September, which still runs behind the scenes.

What changed on 8 October

The early end of the range now comes from two quicker signals: how fast the sister gauge on Clanfield Brook is rising, and how much rain has fallen on our own catchment over the last two days, weighted by how long it takes to reach the brook. Whichever of the two is higher sets the early figure. The low end is still the cautious model, and red still needs the cautious model to agree.

Clanfield Brook is a neighbouring stream, not ours: no water from it reaches Bampton. It sits in the same weather and usually reacts a few hours sooner, which is what makes it a useful warning.

Past winters replayed, with the real delay in when gauge readings reach us
	New model (shown now)	Previous model
Floods warned at least 4 hours ahead	18 of 18	15 of 18
False orange days	48	97
False red days	3	5
Storm Bert, orange	25 h ahead	19 h ahead
Storm Bert, red	5 h ahead	8 h ahead

It warns of every flood in the record, with half as many false orange days. The price is red: on Storm Bert it turned red 5 hours ahead, where the previous model managed 8. The orange before it came 25 hours ahead.

The previous model still runs behind the scenes, and every model is scored on every 15-minute tick, so if the new one does worse on live weather we will see it and can switch back the same day.

Honest caveat: the Environment Agency has been publishing far fewer Clanfield readings since mid-September. The replay above already includes the case where those readings arrive too late to use at all; then the new model leans on the rain signal and still warned of 17 of the 18 floods.

This summary was made with an AI tool. Corrections are marked in the transcript.

5:11 · MP3, 2.5 MB

Read the transcript (6 corrections marked)

Imagine trying to navigate this really complex maze, completely blindfolded.
Oh, that sounds awful.
Yeah, and your only guide is this voice in your ear just reading a weather forecast.
You would definitely bump into a few walls.
Oh, absolutely.
So welcome to this deep dive into the Bampton Flood Monitor.
Today, we are digging into how a tiny village is building its own early warning system.
But first, we really owe you a warm apology.
Right, yeah. The obvious elephant in the room.
Exactly.
There is a deep irony in two very American voices sitting here
analyzing local water levels for a quintessentially English village in Oxfordshire.
I mean, the irony is real, but the lessons here really do apply anywhere.
Before we get into the data, though, we have to be super clear about what this site actually is.
Yeah, it's not official, right?
No, not at all.
bamptonflood.uk is a totally unofficial community website,
just run by a resident who lives right by the brook.
For official flood warnings, you absolutely must use the Environment Agency on gov.uk.
Right. And obviously, if there's ever a danger to life, dial 999 immediately.
Exactly. Never rely on an experimental community site for an emergency.
So looking at their history, you really see how steep the learning curve is here.
I mean, the site only launched in January 2026.
Yeah, it is very young.
And the historical data they're working with, it only goes back to late 2019.
Which is wild because that covers only two floods that actually reached property risk levels.
Just two. You have Storm Bert in November 2024 and another flood on 24 September 2024.
That is really not a lot of data to train a system on.
No, it's really not. But the monitor is trying.
Every 15 minutes, it checks the Environment Agency gauge on Shill Brook
right by the Mill Lane footbridge.
Right. It grabs that current water level.
Yeah.
But to predict what is coming, the system has to rely heavily on standard weather forecasts.
Which brings us back to that blindfolded maze analogy.
I mean, if you have ever relied on a weather app that promised clear skies
while you're literally standing in a downpour, you know the vulnerability.
Oh, yeah. The weather forecast is the site's biggest factor,
but it is also absolutely its weakest link.
And the Storm Bert versus Storm Dennis story really proves that, doesn't it?
It really does.
So on the Sunday lunchtime before Storm Bert hit,[Correction: Sunday 24 November 2024 was during Storm Bert, not before it — rain was already falling. The floodwater reached Bampton overnight into Monday 25 November; in the replay, the early model was already red from midnight.]
the local data looked practically identical to Storm Dennis back in February 2020.
Wait, identical? Like the river levels and everything?
Yeah, river levels, ground saturation, all of it.
But Bert went on to flood Bampton and Dennis didn't.
Wow. So nobody could tell the difference.
Not until the extra rain actually fell over the hills.
The models just couldn't predict precisely where and how hard it would hit.
So to fix that blind spot, they actually updated the system on 17 September 2026.[Correction: the two models do not fix that blind spot. They were added to cut false red alarms (56 to 5 days in a six-winter replay) while keeping the early orange. When the rain forecast is wrong, both models are wrong — Storm Dennis is the example.]
They started running two different prediction models side by side.
Right. Because one just wasn't enough.
So you get this early but jumpy model and then a cautious model.
Okay. So what does the jumpy one do? Just freak out at every rain cloud.
Basically, yeah. It gives you a heads up.
It triggers the green, yellow, or orange alerts to buy you time.[Correction: the colours are shown on the site; the site does not yet send alerts. SMS and email alerts are being built and are not live.]
And the cautious model.
Think of it like a bouncer at a club checking IDs twice.
It actually waits to see physical movement on the river gauge to filter out passing showers.[Correction: the cautious model does not wait for the gauge to move. It uses the same four inputs (brook level, rain in the last 24 hours, the rain forecast, the wetness index) with a much gentler rain multiplier (×0.1 in most wetness bands) and no wetness bonus, so its figure stays close to the measured level until a lot of rain is forecast.]
A red alert only gets past the bouncer if both models agree.
But I do want to push back on something here.
When you look at the site's output, it doesn't give a definitive prediction.
Right. It gives a range.
Yeah. Like it will say expected peak, 1.75 to 2.29 meters.
If I'm a resident, I want an exact number.
Honestly, offering a wide range feels a bit like a cop out.
Why not just give the average?
Well, because an average creates a false sense of certainty.
Giving that wide range is a radically honest admission of, well, we are not totally sure.
Okay. That makes sense.
It is all about balancing alertness against alarm.
They would much rather trigger a deliberate, potentially wrong orange heads up than have[Correction: Storm Bert did flood homes in Bampton overnight into Monday 25 Nov 2024. The site did not exist yet (launched January 2026); Bert is used here as the replay case, showing what the site would have said.]
residents wake up to flooded homes, which tragically happened during Storm Bert.
Right. But they also don't want to cry wolf with red alerts.
Exactly. If they trigger false red alarms all the time,
people get fatigued and just stop paying attention.
Which brings us to the future of the site.
They are hoping to bypass those generic, gridded weather forecasts entirely[Correction: upstream rain sensors are a hope, not a plan with a date, and none are installed. They would measure rain that has already fallen on the hills; the forecast of rain still to come would still be needed, so they would not bypass forecasts entirely.]
and install new rain sensors on higher ground upstream, right?
Yeah. To measure actual physical rain falling on the hills before it hits the brook.
But we should be clear. This is strictly a plan right now.
It is not actually in place yet.
Right. It's a hope for the future.
But it really shows the whole community ethos behind this project.
This resident openly publishes all the system's prediction errors on the site.
Oh, that's refreshing. So there is no blame when things go wrong.
None at all. Just a relentless focus on fixing it.
It is just an understanding that residents, volunteers, and local councils are all doing
their absolute best with the tools they have.
Which leaves you with a really fascinating question to ponder.
As weather gets more unpredictable, you just have to wonder,
will our massive, multi-million dollar national forecasting models
eventually be outpaced by these sort of hyper-local,
crowdsourced networks built by neighbors in their own backyards?

Transcribed locally with faster-whisper. Names the transcription misheard (Shill Brook, Storm Bert, Storm Dennis) were corrected in the text; nothing else was reworded.

Try it yourself

Think you could predict the brook better?

Move the sliders, replay real storms on Shill Brook, and see if you can beat our model.

1. What a prediction is

Every 15 minutes the site takes a fresh reading and estimates how high Shill Brook at Bampton will peak in the next day or so, then turns that number into a colour. It is a guess about the future peak, not the level right now — the actual level is shown separately.

Colour	Predicted peak	Plain meaning
Green	under 1.06 m	Normal
Yellow	1.06 – 1.40 m	Brook up, nothing to do
Orange	1.40 – 1.78 m	Flooding of low-lying land possible
Red	1.78 m and over	Property flooding possible

2. What every model looks at

Both models use the same four ingredients at every tick: the brook level now (Environment Agency gauge), rain in the last 24 hours, the rain forecast for the next 24 and 48 hours, and a wetness index — a running score of recent rainfall, recalculated once a day at midnight, used as a stand-in for how soaked the ground is. It is not a soil-moisture measurement.

The sum is simple: predicted peak = brook level now + (rain × a multiplier), capped at 2.50 m. The only difference between the models is the size of the multiplier, which depends on which wetness band the ground is in.

3. The two models

The live model (v3.1) is what the site has run all year. The wetter the ground, the harder it multiplies the rain (×0.5 when dry up to ×2.5 when very wet), plus a wetness bonus of up to 25 cm. Its strength is that it is early: it went red 21 hours before Storm Bert’s floodwater arrived. Its weakness is that it is jumpy: the wetness index changes only at midnight, so the prediction can leap in one step, and over six winters it called red on about 56 days when no red flood came.

The cautious model (“Q-A”) uses the same ingredients and the same sum with a much gentler multiplier (×0.1 in most wetness bands, ×1.1 in one) and no bonus. It mostly stays close to the real brook level unless a lot of rain is on the way. Its strength is that it is rarely wrong: about 5 false red days in six winters, and 1 in the 2025/26 winter. Its weakness is that it is later: red about 8 hours before Storm Bert’s floodwater, and only 2 hours ahead of the September 2024 flood.

Honest caveat: its settings were chosen while looking at past winters, so they may be partly tuned to those storms. Live weather is the real test.

4. The agreement rule

Neither model is good at everything, so each gets a job. Green, yellow and orange come from the live model — it is early, so it gives the heads-up. Red only shows when both models say red. If the live model says red but the cautious one does not agree, the site shows orange instead. Nobody has to pick a river level at which to “swap models”; the cautious model acts as a second opinion before anything turns red.

Six winters of real data (2019/20 – 2024/25), replayed
	Live model alone	With the agreement rule
False red days	56	5
False orange days	97	97 (unchanged)
Real floods warned	all 16	all 16
Storm Bert	red 21 h ahead	orange 19 h ahead, then red 8 h ahead

Winter 2025/26 (no real red): false red days 19 → 1, false orange days 40 → 40, all 4 real oranges warned. The rule was tested after the winter, so that winter no longer counts as unseen.

5. What number goes on the screen

The colour is only half the box. Showing one model’s number beside a colour decided by both would contradict itself — “Orange” next to “2.00 m” makes no sense. So the front page shows a range, from the cautious model’s figure to the live model’s, and the colour follows the agreement rule. The low end is the cautious estimate; the high end is the early, pessimistic one. A narrow range means the two models agree. A wide range means real uncertainty, and residents can see it.

Worked examples, real data
When	Live	Cautious	Box would show	What happened
Storm Bert, Sun 24 Nov 2024, 06:45	2.29 m (red)	1.75 m (orange)	ORANGE — 1.75–2.29 m	At 12:45 Q-A reached 1.78 m, both said red, the box turned RED — 1.78–2.32 m. The brook peaked at 2.17 m, inside that range.
False alarm, 17 Nov 2022	2.46 m (red)	1.00 m (green)	ORANGE — 1.00–2.46 m	The brook never passed 0.96 m. The very wide range honestly signals “we’re not sure”. Under the live model alone this was a red.
Storm Dennis, 16 Feb 2020, 15:00	2.34 m (red)	1.87 m (red)	RED — 1.87–2.34 m	The brook peaked at 1.39 m. One of the 5 false reds the rule does not remove: when the forecast is wrong, both models are wrong.

6. What you see

The same four colours, from the same data, updated every 15 minutes; a range instead of a single number; orange earlier and more often — the heads-up that rain is coming onto wet ground and this could build, which will be wrong fairly often and says so on the box; and red far less often, but when it shows, both models agree. It means “take this seriously now.” You will not see model names on the front page, just the two ends of the range.

Limits worth knowing

Weather forecasts can be wrong. Every model leans on them. At lunchtime on Storm Bert, the data looked almost identical to Storm Dennis (Feb 2020), which never flooded. No model can separate those two until the extra rain actually falls.
There is no rain gauge upstream of Bampton, so rainfall comes from gridded weather data.
Only two real red floods exist in the six-winter record (Storm Bert, 24 Sep 2024), so red performance rests on very little evidence.
Winter 2025/26 has now been used for testing and can’t be used to tune anything further.

Full daily record

Every day since monitoring began: the gauge at 6am, noon and 6pm, the day's high and low, rainfall and soil wetness.

The floods the model learned from

Every recorded flood event at Bampton: its peak, the rain behind it, and how long the brook took to respond.

Total Events

Avg Lag Time

N/A

Avg Rise Rate

N/A

Max Level

N/A

Worst Event

All flood events

No flood events recorded yet.