How accurate is a forecast 1, 3, 5 and 7 days ahead?

Measured against the weather that actually happened, a one-day forecast is accurate to about 0.79 degrees and a seven-day forecast to about 1.93 degrees - roughly 2.4 times the error. In practice a 1-3 day forecast is reliable enough to plan around, 4-7 days is a useful steer that will shift, and beyond about ten days a single deterministic forecast tells you little more than the seasonal average for that place. The more useful signal at longer range is whether the models agree with each other.

Not in the way they are presented. We measure model error out to seven days, where it has already roughly grown 2.4-fold from day one. Beyond ten days the atmosphere's own predictability limit dominates, and a 14-day daily temperature is closer to a climatological guess than a forecast. Apps still show them because people click on them. Treat anything past ten days as a trend, not a plan.

Small errors in the starting conditions grow as the simulation runs forward - the atmosphere is chaotic, so two nearly identical starting states diverge over time. That is also why running several models is informative: when independent models still agree at day six, the pattern is genuinely locked in. When they diverge, the uncertainty is real rather than a modelling failure.

For a sample of cities in each market we pull what each model predicted for the same valid hour 1, 2, 3, 5 and 7 days in advance, then grade every one of those predictions against ERA5 reanalysis - the recorded weather. The figures above are mean absolute error in degrees. Where a model does not publish that far ahead we show a dash rather than filling the gap.