Reading the page
The big tile is the official OURCs flag, from
ourcs.co.uk. The cards give
the chance it changes today or by the end of tomorrow, and where the differential is heading. In the chart, the
black line is the measured differential; each thin purple line is one possible future from a different rain
forecast, the bold line is their average, and the shading is the likely range (rain and model error together;
outer 80%, inner 50%). The chart below it gives the chance of each band over the next 10 days.
The flag and the differential
The coloured bands are ranges of the differential that roughly match the flags, but OURCs sets the flag by hand,
so the two agree only about 80–87% of the time. Flags go up within an hour or two of the differential
crossing a line, come down later once it is clearly below, and almost only change between 8am and 8pm; the flag
card's chances come from a model fitted to the OURCs flag history. For the first three days the forecast moves in
held levels and daytime jumps, as lock keepers adjust the weirs, rather than a smooth line. Black (Isis) is
forecast from the Iffley Lock level (Black from about 0.42 m, back to Red below 0.22 m), and its card
appears only when Black is close.
How good is it?
Treat it as
experimental. Checked on years it was not trained on, cards saying 80% or more came
true about 9 times in 10, and 60–80% about 6 to 7 times in 10. Under each chart you can compare any recent
forecast with what the river then did. The model cannot rise without rain, but it can overshoot after heavy rain
and still misses some weir changes. The summary sentence is written automatically from the average forecast.
Data and updates
River levels and rain from the Environment Agency; rain forecasts from ECMWF AIFS (50 members) via Open-Meteo.
The model is a physics-based rainfall–runoff model whose starting state is set by a small neural network,
averaged over five trained copies. It runs every 15 minutes, from the newest gauge readings. If a stretch's gauges go offline, its differential is estimated from
the neighbouring stretch (the Isis from Godstow, and the other way round) until they are back, with a wider
likely range; after five days offline the estimate is marked low confidence. With both offline, the forecast is
marked stale. Write-ups are on the
substack; bug reports and
ideas are welcome on
GitHub.