WeatherNext 3 Precipitation Forecast: Reading Rain Hour by Hour
Precipitation is the variable WeatherNext 3 changed most. Instead of inheriting rain from processed analyses, the model learns it from satellite-derived estimates at 0.1 degree, and Google's own evaluation reports a 60% improvement in precipitation skill against the IMERG reference. This page explains where that signal comes from, how to read it in the hourly output, and which caveats the paper itself sets out. For the wider variable picture, see the model overview.
7-day free trial, then $19/mo · cancel anytime · compare plans
The three numbers that matter
Each figure comes from Google's published material; the commentary is ours.
0.1° Precipitation is learned at 0.1 degree on hourly steps, matching the resolution class of the sharpest single-level variables in the model.
Precipitation is learned at 0.1 degree on hourly steps, matching the resolution class of the sharpest single-level variables in the model.
60% Reported improvement in precipitation skill against the IMERG reference in Google's own evaluation - a relative figure, not a guarantee for your region.
Reported improvement in precipitation skill against the IMERG reference in Google's own evaluation - a relative figure, not a guarantee for your region.
1 hour New global precipitation fields arrive hourly rather than on a six-hourly cycle, which is what makes rain timing usable for short decisions.
New global precipitation fields arrive hourly rather than on a six-hourly cycle, which is what makes rain timing usable for short decisions.
Where the precipitation signal comes from
Rain is the hardest variable to keep honest, because ground truth is patchy. This generation changes the inputs rather than only the architecture.
Geostationary observations
The model reads raw satellite mosaics - infrared and visible bands - instead of relying only on pre-processed analysis fields.
Satellite rain estimates
Precipitation is learned from satellite-derived products, including IMERG-based estimates, which is why the reference for evaluation is a satellite product rather than a gauge network.
PARDIG
The paper describes a second precipitation product the team calls PARDIG, which its internal checks rank closer to independent observations.
Ensemble members
Rain is produced as part of a 64-member ensemble, so the useful output is a range of outcomes rather than a single number.
How to read precipitation in the output
Four habits that turn a precipitation field into a decision that survives contact with reality.
Know which quantity you are looking at
Rain rate and accumulation are different things. A 0.6 mm/h hour and a 0.6 mm total look identical on a chart and mean different things for a commute or a field.
Use the spread, not one run
Members that agree suggest a settled outcome; members that scatter describe a probability. Treat a single member as an example, never as the forecast.
Check the timing, not just the amount
Hourly output exists so the arrival window can be narrowed. The difference between rain at 14:00 and rain at 20:00 is usually the decision.
Compare with what you can see
Radar and local gauges remain the honest check for the next couple of hours. Use them to calibrate your reading of the model, not to replace it.
What to keep in view
The paper is unusually explicit about the edges of its precipitation result. These are the ones that matter in practice.
Partly self-referential verification
Precipitation is verified in part against products the model also trained on. The paper states this; anyone quoting the 60% figure should state it too.
Regional variation is real
Reported gains are strongest where forecasting was historically weakest. Where your local baseline is already good, expect a smaller difference.
Extremes deserve their own check
Convective downpours and orographic rain sit at the edge of every global model. Judge those cases separately instead of trusting an average.
Warnings stay official
Nothing in a model output replaces a national meteorological service warning. Use the forecast to plan, and the agency to decide when safety is involved.
WeatherNext 3 precipitation questions
Short answers about the rain variable, the reported gains and how far to trust them.
Watch the rain hour by hour
Precipitation fields, hourly steps and ensemble context in one workspace. Seven-day free trial, cancel anytime. Background: the paper notes and the accuracy guide.
Related guides
The pages that go with this one, chosen by what each one answers.
