What should I do when the models disagree?

Wind models are mathematical simulations of the atmosphere. They use vast amounts of data and complex equations to predict wind. Different models use different equations, data inputs and ways of representing the atmosphere, so they can give different forecasts for the same place and time. ECMWF IFS, GFS and ICON are three distinct models, each with its own approach. When they show different wind speeds or directions, that reflects real uncertainty in forecasting. It is not a fault, and no single model is always right. The instrument shows you the disagreement so you can take it into account.

The instrument shows disagreement in several places. The chip at the top left of the map reads AGREE or SPLIT, with the number of models, such as 3/3. It reads SPLIT when the agreement score of the three deterministic models is under 60 out of 100. The agreement spine in the inspector holds five ticks, and two of them are about spread. Model spread says whether the deterministic models agree or split for that hour. Ensemble IQR gives the P10 to P90 range of the ensemble, marked tight or wide, and says when the primary forecast and the ensemble median fall on opposite sides of your limit. Hover or tap a tick to read its detail. A wide range or a SPLIT means more uncertainty.

Always compare the forecast to the nearest observation. The observation line shows the latest METAR or buoy report within 150 km, its age and distance, and what the model said for 10 m at the same hour. One reading is a check, not a trend, and a station can be a long way from your site, so a mismatch is a prompt to look closer. A line marked STALE, or older than 90 minutes, tells you less. If no usable station is found, the line says so, and you are checking the models only against each other.

The ensemble shows many possible outcomes rather than one. The exceedance fan draws a band for the P10 to P90 range of the members, a stronger band for P25 to P75, and a line for the median alongside the primary forecast. A narrow band suggests more agreement, and a wide band suggests less. Pay attention to the share of members at or above your Limit, shown as P(>limit). It is a fraction of members, not a calibrated probability, and the instrument flags EXCEED when it reaches 40%. Read the exceedance fan explains the chart.

When the models split or the ensemble band is wide, it is prudent to widen your own margin. If your document gives a limit, you might decide to work to a lower number for that period. For example, with a limit of 10 m/s, you could choose to treat 8 m/s as your working limit. This buys a buffer against forecast error. The instrument does not make that choice for you. It uses the limit you type, and the decision to adjust it rests with the person responsible on site, on the strength of their own risk assessment.

Models are re-run at regular intervals, so newer data can narrow the gap or bring the models closer together. Check the fetched time in the provenance note at the foot of the inspector, and the run times of the models, to see how fresh the data is. A more recent forecast, even if it still shows some disagreement, is generally better than an older one. You can find more in How fresh is the data.

When models disagree, do not simply average the wind speeds into one number. Averaging hides the uncertainty. If one model gives 5 m/s and another 15 m/s, a 10 m/s average does not describe the risk. The agreement spine is labelled EVIDENCE, NOT CONSENSUS for this reason. Look at the full range, and at the odds of crossing your limit. A SPLIT chip or a wide ensemble band means a single deterministic value is not enough to decide on.

It is good practice to keep what you saw when you decided, especially when the models disagreed. The Freeze record button in the instrument makes a time-stamped copy of the forecast you were viewing, with the place, hour, height, limit, numbers, member counts and sources. It does not update afterwards, so it answers what the forecast said at that moment. That is useful for later reviews. See Freeze a record for the steps.