Have you ever opened two weather apps on the same morning and been handed two different tomorrows? One puts the rain in at noon, the other clears you for a dry evening, and both deliver the verdict in the same clean type, with the same flat confidence, as if they were reading a thermometer instead of running a simulation.
Neither app is lying to you. They are reading different models, and the gap between them is the most useful thing on the screen once you know what it means.
What Is A Forecast, Exactly?
A forecast is the output of a simulation. Physics equations are solved on a three-dimensional grid, starting from an imperfect snapshot of the atmosphere and stepped forward in time, hour by hour.
Nobody has measured tomorrow. The number on your screen came out of a numerical weather prediction model — a set of equations for fluid motion, radiation, moisture, and heat, solved over a grid that wraps the planet from the surface to the stratosphere.
That process happens in two stages. First comes data assimilation, which blends millions of observations — surface stations, ships and buoys, aircraft sensors, radiosondes released twice a day at 00Z and 12Z, and above all satellite radiances, which supply the overwhelming majority of the information over open ocean — into a single best-guess analysis of the atmosphere right now.
Then the model integrates forward, recomputing every grid cell from its neighbors in increments of a few minutes of simulated time. Whatever was wrong in the starting snapshot travels with it and grows.
That analysis is where the disagreement is born. Two forecast centers, running two assimilation systems on two slightly different sets of observations, produce two slightly different versions of right now — and a slightly different right now becomes a genuinely different Thursday.
Which Models Is Your App Actually Reading?
Almost every consumer app is downstream of two global models: NOAA's GFS and the ECMWF, sometimes wrapped in a proprietary blend. Very few apps run any physics of their own.
The logo on the app is a design decision. The forecast underneath it is nearly always one of two systems, or a statistical mix of them.
The GFS (American)
The Global Forecast System is run by NOAA's National Centers for Environmental Prediction four times a day, at 00, 06, 12 and 18 UTC. It carries roughly 13 km horizontal resolution through the first ten days and 127 vertical levels, and it publishes into the public domain.
That last fact explains its reach. Because the GFS is free and openly served, it sits under a large share of the weather apps, dashboards, and widgets you have ever used — and because it runs four times daily, it refreshes twice as often as its European rival.
The ECMWF (European)
The European Centre for Medium-Range Weather Forecasts runs the Integrated Forecasting System from Reading, England, twice a day at 00 and 12 UTC, at roughly 9 km resolution. On the standard medium-range verification metric — 500 hPa geopotential height anomaly correlation — it has led the world for most of the last three decades.
Its reputation was cemented in October 2012, when ECMWF runs showed Hurricane Sandy hooking left into New Jersey roughly a week before landfall while GFS runs still carried the storm harmlessly out to sea. The gap was public, expensive, and directly motivated a decade of American investment in supercomputing and model development.
| Model | Who runs it | Update cadence | Resolution | Strongest at |
|---|---|---|---|---|
| GFS | NOAA / NCEP (United States) | Four runs daily (00, 06, 12, 18 UTC) | ~13 km, 127 levels | Frequent refreshes; free, so it is everywhere |
| ECMWF (IFS) | ECMWF (Reading, England) | Two runs daily (00, 12 UTC) | ~9 km | Medium-range pattern, days three to seven |
| GEFS | NOAA / NCEP | Four runs daily | ~25 km, 31 members | Confidence and scenario spread |
| ECMWF ENS | ECMWF | Two runs daily | 51 members | Probabilities in the medium range |
| HRRR | NOAA / NCEP | Hourly | 3 km, 18–48 hours | Storms, squall lines, the next six hours |
The table is the short version. The practical takeaway is simpler: when your two apps disagree, there is a good chance you are holding a GFS app in one hand and a Euro-fed app in the other, and the disagreement between them is a real measurement of how uncertain tomorrow actually is.
Why Do Two Models Drift Apart After Day Three?
Small errors in the starting analysis double roughly every day or two. By day three they have grown to the size of real weather systems, so two nearly identical starting points produce two different forecasts.
Edward Lorenz found this in 1963, when a rounding difference in the fourth decimal place sent a simulated atmosphere down a completely different path. The atmosphere amplifies small errors relentlessly, and that property has a name: chaos.
Three things drive the divergence, and they compound. The first is initial-condition error, which no observing network can eliminate, because the oceans and the upper atmosphere are sampled thinly and the model has to interpolate across the gaps.
The second is resolution. A 9 km grid and a 13 km grid do not see the same mountain ridge, the same lake breeze, or the same thunderstorm — and terrain is where a great deal of weather is actually made.
The third, and the least visible to readers, is parameterization. Anything smaller than a grid cell — convective updrafts, cloud microphysics, boundary-layer turbulence, how much heat the land gives back at dusk — cannot be resolved, so it is approximated with a scheme, and each center writes its schemes differently.
All of these add up to a predictable decay curve. Inside two days, models nail temperature within a couple of degrees and timing within an hour or two; through days three to five the pattern holds but the details go soft; by days six to ten you have a pattern and nothing more; past day ten, climatology is usually as honest as the model.
What Does An Ensemble Run Actually Do?
It runs the same model dozens of times from slightly perturbed starting points. The spread between members is the forecast's confidence interval — tight clustering means high confidence, wide scatter means a fork.
If the starting analysis is uncertain, the honest response is to quantify the uncertainty rather than hide it. So the centers nudge the initial conditions and run the model again, and again, and again.
NOAA's GEFS runs 31 members on a coarser grid four times a day. ECMWF's ENS runs 51 members twice a day, and both produce a distribution rather than a line.
The classic visualization is the spaghetti plot, which draws one 500 hPa height contour from every member on a single map. When the strands run together, the large-scale pattern is locked in and you can plan; when they fray into a bush, the atmosphere is standing at a fork and no app should be telling you a single number with a straight face.
This is also where probability comes from. That 40% on your screen is post-processed ensemble output, not a forecaster's hunch — and it means something more specific than most people assume, which we unpack in what a 40% chance of rain actually means.
One caution about ensembles: the mean is not the forecast. Averaging 51 members smooths a deep, sharp low into a shallow, blurry one, so the ensemble mean is excellent for pattern and systematically bad at extremes — read the members, not just their average.
Which Model Owns The Next Six Hours?
Neither global model does. Inside roughly twelve hours, a convection-allowing model like the 3 km HRRR — refreshed hourly and initialized with radar — beats both on timing and storm structure.
The High-Resolution Rapid Refresh runs on a 3 km grid, updates every single hour, and reaches out 18 hours (48 at the main synoptic runs). Critically, it assimilates radar reflectivity, which means it can initialize storms that already exist rather than trying to invent them from scratch.
At 3 km the model is convection-allowing, meaning it can represent individual thunderstorm updrafts instead of approximating them with a scheme. That is why a good short-range model can put a squall line on your street at 4:15 in the afternoon and be roughly right, while a global model can only tell you it will be an unsettled day.
There is one more layer most readers never hear about. NOAA's National Blend of Models is a statistically calibrated combination of many models, bias-corrected against station history — and your local National Weather Service forecast starts from the Blend before a human forecaster edits it.
Which means the official forecast is already a multi-model consensus with a person on top of it. And for the next sixty minutes, no model beats the radar loop itself, which is why learning to read radar is still the highest-leverage weather skill a person can pick up.
How Should You Read Disagreement?
Read agreement as confidence and disagreement as a warning to stay flexible. When apps split inside three days the number is uncertain; when they split past five, the whole pattern is.
A split screen is data, and it can be translated into a decision. Here is how we translate it:
- Check the clock before you check the model. GFS output lands a few hours after each run and the ECMWF later still, so a 6 a.m. disagreement is sometimes just one app holding a stale run while the other has refreshed.
- Separate a timing disagreement from an amount disagreement. If both models have rain and only the hour differs, you have rain and you should carry the shell; if one has rain and the other has none, that is a genuine fork in the atmosphere.
- Watch run-to-run consistency, not a single run. A model that has shown the same solution for four consecutive runs is telling you something that a model flipping every twelve hours is not.
- Weight by lead time. Inside twelve hours, trust the high-resolution short-range model and the radar; from days three to seven, trust the ensembles and the ECMWF's read on the pattern.
- Convert spread into a plan rather than a number. Uncertainty is an argument for layers, a backup window, and a bag that covers both branches — the logic we lay out in packing by forecast.
All of the above adds up to one habit. Stop asking which app is right and start asking how far apart they are, because the distance between them is the forecast's own estimate of how much it knows.
The practical rule. Inside 48 hours, trust the short-range model and the radar over any global run. Past five days, stop reading the number and read the spread.
Where Do The AI Models Fit?
Since 2023, a genuinely new class of forecast model has arrived: systems trained on decades of reanalysis data rather than solving the physics from first principles. DeepMind's GraphCast, Huawei's Pangu-Weather, and ECMWF's own AIFS all belong to this family, and ECMWF now runs an AI system operationally alongside its physics model.
The speed difference is not incremental. A ten-day global forecast that takes a supercomputer hours can be produced by a trained model in about a minute on a single machine, and on standard verification scores these systems have matched or beaten the physics models on many medium-range variables.
Two caveats are worth holding onto. They are trained on the past, so their behavior in genuinely unprecedented conditions is an open question, and they still depend on a physics-based analysis to initialize — which means they inherit exactly the same starting-error problem that makes day five hard for everyone.
Expect your app to start blending them in quietly, without telling you. The disagreement you see next winter may well be a physics model arguing with a neural network.
How We Read The Split At Vesper
Most weather apps make the choice on your behalf and hide it. Apple Weather gives you one number, Carrot dresses the same number in personality, and The Weather Channel gives you the number with an ad beside it.
We think the hidden choice is the actual problem. When the models disagree, we would rather say so — because "the models split on Thursday afternoon" is a fact you can plan around, and a lone confident number that quietly resolves the fork for you is not.
So the brief names the dominant story of the day, names the uncertainty where it exists, and translates both into a decision: a garment, an hour, a shoot window. That method is written down in how we write a brief, and the argument behind it in weather worth reading.
It matters most for light. Cloud cover is among the least skillful fields any model produces, which is precisely why we built Sunset Verify instead of trusting a five-day cloud icon — and why the rest of the journal spends so much time on the difference between what the model says and what the sky does.
Read the brief tomorrow. It will tell you when we are sure, and it will tell you when we are not.
Common Questions
Which weather model is the most accurate?
The ECMWF has led medium-range verification scores for decades and is the usual benchmark, which is why forecasters call it the gold standard. But accuracy is scale-dependent: for the next six hours, a 3 km convection-allowing model like the HRRR — refreshed hourly and fed by radar — will beat both global models on storm timing and structure.
Why did my forecast change overnight?
Because a new model run landed. The GFS updates four times a day and the ECMWF twice, and each run ingests a fresh batch of observations from satellites, balloons, aircraft, and surface stations. A forecast is the best available guess at the hour it was made, not a commitment — and a changed forecast usually means the model is working correctly.
What does an ensemble forecast actually tell me?
It tells you how confident the atmosphere is willing to let the model be. The same model is run 31 or 51 times from slightly different starting points; tight clustering between members means the pattern is locked and you can plan on it, while wide scatter means the atmosphere sits at a genuine fork and you should keep both branches in the plan.
How far out is a weather forecast actually useful?
Roughly: hours one through twenty-four for precise timing, days one through three for specifics like temperature and rain windows, days four through seven for the pattern, and beyond that for tendency only. Past about ten days, climatology — what that date usually does in that place — is generally as honest as any model output.
Should I trust the app that agrees with my plans?
No. Picking the friendlier forecast is confirmation bias wearing a weather icon, and the atmosphere is indifferent to your Saturday. When two credible apps disagree, the correct reading is that confidence is genuinely low, and the right response is a flexible plan — a shell in the bag, a backup window — rather than the number you prefer.