Why do weather forecasts disagree?
Different models, imperfect starting data and a chaotic atmosphere mean forecasts diverge — and that disagreement is your best guide to how much to trust them.
Weather forecasts disagree because there is no single forecast. Different national agencies run different numerical models — ECMWF, GFS, ICON and Australia's own ACCESS — each with its own physics, resolution and starting snapshot of the atmosphere. Because the atmosphere is chaotic, tiny differences at the start grow over time, so the models diverge — especially several days out. That disagreement is not error so much as a measure of genuine uncertainty.
The core reason: there is more than one model
When two weather apps show different forecasts for Saturday, they are usually just showing different models, or the same model from a different run. ECMWF (European), GFS (American), ICON (German) and ACCESS (Australian) are separate systems built and maintained by separate agencies. They are all solving the same physics, but they make different choices about how — and those choices lead to different answers. One app leaning on GFS and another on ECMWF can genuinely disagree while both are working perfectly.
Initial conditions and the butterfly effect
Every forecast begins with a snapshot of the atmosphere assembled from satellites, balloons, aircraft, ships and ground stations. That snapshot is never perfect — there are always gaps, especially over oceans and remote regions. In 1963 Edward Lorenz discovered that tiny differences in a model's starting state grow exponentially, the origin of the "butterfly effect". So even two runs of the same model, started from slightly different observations, will eventually diverge. Multiply that by four different models and you have a recipe for disagreement that grows with forecast range.
Different physics and resolution
Models also differ in their internals:
- Resolution. ECMWF runs at roughly 9 km, ACCESS-G and GFS nearer 12–13 km. A finer grid resolves fronts, coastlines and terrain more sharply, which changes the forecast.
- Parametrisation. Processes too small to simulate directly — cloud formation, convection, turbulence — are approximated with "parametrisation schemes". Each model does this differently, and those choices particularly affect rainfall and thunderstorms.
- Data assimilation. How each model blends millions of observations into its starting snapshot varies, so even the moment of "now" differs slightly between models.
None of these differences is a mistake. They are legitimate, expert engineering choices — which is exactly why comparing models is informative rather than confusing.
Ensembles: models that disagree with themselves on purpose
Forecasters lean into uncertainty using ensembles. Instead of one run, an ensemble runs the model dozens of times with slightly perturbed starting conditions. If all the members land in roughly the same place, confidence is high; if they scatter, the atmosphere is in an uncertain state. ECMWF, NOAA and the Bureau all run ensembles for exactly this reason. The spread of an ensemble is, in effect, a scientific measurement of how much to trust the forecast.
The same model, a different run
Disagreement isn't only between models. The same model run six hours apart can shift its story, because each run assimilates a fresh batch of observations. If you check a forecast at breakfast and again after dinner and it has changed, that usually means new data nudged the model — not that anyone got it wrong. This "run-to-run consistency" is itself a confidence signal: a forecast that holds steady across several consecutive runs is more trustworthy than one that lurches around from run to run.
Why disagreement grows with distance
Model spread is small in the near term and widens with time. At day one, ECMWF, GFS, ICON and ACCESS will usually be close, because errors haven't had time to amplify. By day seven or eight, they can paint quite different pictures for the same city. This is why a disagreement two days out is a genuine warning sign worth heeding, while a disagreement nine days out is completely normal and rarely worth losing sleep over. Judging disagreement always means asking: how far ahead is this?
Why disagreement is useful, not annoying
Here is the mindset shift. Model disagreement is not noise to be ignored — it is one of the most valuable pieces of information you can have. When ECMWF, GFS, ICON and BOM's ACCESS-based guidance all agree that a cold front crosses Perth on Friday, you can plan with real confidence. When they split — one wet, one dry, one a day late — the honest conclusion is that Friday is genuinely uncertain, and you should keep your options open and re-check tomorrow.
This is the whole idea behind checkweather.io. Rather than picking one model and pretending it is the truth, it lines up ECMWF, GFS and ICON side by side, measures how much they agree, and turns that spread into a plain-English forecast-confidence score — then grades past forecasts on an accuracy scoreboard against what actually happened. No black boxes: you see the disagreement and what it means. For the official Australian picture, it points you to BOM.
How to use disagreeing forecasts
- Look for agreement, not a favourite. Don't pick the app you like; check whether the major models concur.
- Weight the near term. Disagreement at day two is a real red flag; disagreement at day nine is normal and expected.
- Defer to BOM for decisions. For warnings, cyclones and fire danger, the Bureau's official product is the authority, whatever a model layer shows.
- Re-check after new runs. Each run ingests fresh data, and disagreement often resolves as an event approaches.
The bottom line
Forecasts disagree because they come from different models, started from imperfect snapshots of a chaotic atmosphere, using different physics. That is not a flaw in weather forecasting — it is weather forecasting being honest about uncertainty. Read the disagreement, and it becomes your best guide to how much to trust the forecast in the first place.