🔍
Forecast confidence is a measure of how much the world's leading weather models agree about what is going to happen. When they agree closely, the forecast is far more likely to hold, and confidence is high. When they diverge, some predicting a dry afternoon, others a thunderstorm, confidence is low, and the forecast is more likely to change. Our confidence score turns that agreement into a single number from roughly 5 to 98, so you can see at a glance how much to trust the days ahead.
Modern forecasting does not run a weather model just once. Because the atmosphere is chaotic, tiny differences in today's starting conditions grow into large differences a week out, forecasters run the same model many times, each from a slightly different but equally plausible starting point. This collection of runs is called an ensemble, and each individual run is a member. Our confidence score is built on the ensemble from Google DeepMind's WeatherNext 2 model, which produces 64 members for every forecast.
Think of it as asking 64 well-informed experts the same question. If all 64 say tomorrow's high will be near 22°C, you can be confident. If their answers scatter from 16°C to 28°C, the honest conclusion is that tomorrow is genuinely uncertain, and a good forecast should tell you that rather than hide it behind a single confident-looking number.
The technical heart of the score is the spread of the ensemble, how far apart the members are, measured as the standard deviation across all 64 of them at each point in time. A tight spread (members clustered within a degree or so) means strong agreement and high confidence. A wide spread means disagreement and low confidence. We compute the spread for the near-term forecast, translate it into a 0–100 style score, and widen our expectations honestly as the forecast reaches further into the future, because even the best models become less certain with distance.
Confidence measures model agreement, not certainty of being correct. It is possible, though less common, for all the models to agree and still be wrong together, because they can share the same blind spot. So a high confidence score means "the forecast is unusually predictable right now," not "there is a 90% chance this exact temperature occurs." We think that distinction matters, and we would rather show you an honest measure of agreement than a falsely precise promise of accuracy. This is exactly the kind of understanding that separates a real forecast from a guess dressed up with a number.
Knowing the confidence changes how you use a forecast. When confidence is high, you can plan firmly, book the outdoor event, schedule the harvest, trust the travel window. When it is low, you plan flexibly, keep a backup, delay the irreversible decision, check again tomorrow. A forecast without a confidence signal forces you to treat a rock-solid outlook and a coin-flip the same way. With one, you can act on good information and hedge on uncertain information, which is what good weather decisions have always required.
Confidence comes from how much different forecast models agree. Choose which models to include below, and we'll recompute the score live for your location. Each model has different strengths, see the guide underneath.
Different forecasting centres run different models, and they don't always agree. Pick a few below to see where they line up and where they differ for your location — and get a plain-English guide to which ones suit where you are and what you're doing.
Set a location, pick some models, then press compare.
No single model wins everywhere. Here's a plain-language guide to the ones above, so you can choose a mix that suits what you're forecasting.
Google DeepMind's machine-learning ensemble. Extremely fast and often excellent for the medium range (days 3–10) and for tracking large weather systems. Best when you want a strong all-round default, or an AI "second opinion" alongside physics models.
Widely regarded as the world's most accurate global physics model overall. Best for general-purpose forecasting anywhere, and the strongest single choice for the 1–7 day range in most of the world, especially Europe.
The American global model. Updates frequently and covers the whole planet. Best for North America and for situations where you want frequent refreshes; historically a little noisier than ECMWF but a solid, independent voice in the mix.
Germany's global model, strong over Europe and known for good handling of the short range (1–3 days) and convective setups. Best as a European short-range specialist and a useful cross-check against ECMWF.
The Canadian global model. Performs well at high latitudes and over North America. Best for northern regions and as an extra independent member to widen the ensemble and stress-test agreement.
Confidence appears throughout VWeatherStation: on the forecast page as a full panel with per-day agreement, on individual city pages, and behind our forecasting tools. Wherever you see a confidence score, it is computed the same way, live, from the WeatherNext ensemble spread for that exact location.