How accurate are long range forecasts?

Why forecasts fade with time

The atmosphere is a chaotic system: tiny differences in the starting state grow until they dominate. No network of observations can measure every gust and eddy, so every forecast starts with small errors, and those errors roughly double every couple of days. By about two weeks they have grown to the size of the weather systems themselves.

This is not a failing of computers that more power will fix completely. Better models and observations have pushed useful skill further out, by roughly a day per decade, but the limit set by chaos remains. That is why no responsible forecaster offers a daily forecast for a date a month away.

Ensembles: forecasting the uncertainty

Instead of running a model once, forecast centres run it many times from slightly different starting conditions. The European Centre for Medium-Range Weather Forecasts runs 51 versions for the next 15 days. Where the runs agree, the forecast is solid; where they spread apart, the weather is genuinely undecided.

The chart below is the live ensemble for London. Each faint line is one run. Look at how tightly they bunch for the first few days and how they fan out later. That fan is the uncertainty, made visible, and it is different every time: some weeks the runs agree well into the second week, other weeks they split by day five.

Day 1 to day 15: what holds and what does not

Within the ensemble period, different parts of the forecast fade at different rates. Large-scale patterns, such as a big area of high pressure settling over the UK, are often visible a week or more ahead. Temperature trends hold reasonably well. The timing of a rain band slips first, and the exact track of thunderstorms or showers is rarely predictable more than a day ahead.

So a sensible reading of a ten-day forecast is: trust the first few days in detail, the next few for the general type of weather, and treat the last few as one possibility among several. Our confidence marks do this for you from the spread of the runs, using fixed thresholds published on our data sources page.

Weeks three to six: how the weeks lean

Beyond about two weeks, forecasts switch from predicting weather to predicting the odds. An extended-range system such as ECMWF's EC46 runs around a hundred versions out to 46 days and compares them with the same model's history for that time of year. What comes out is a leaning: week four looks, on balance, milder or wetter than usual.

These leanings carry real skill some of the time, especially when something large is steering the pattern, such as a persistent block of high pressure or a disturbance high above the Arctic. Much of the time, though, the signal is weak, and the most honest answer is close to normal. We say exactly that rather than dressing a weak signal up as a forecast.

Seasonal forecasts: a slight tilt of the odds

Seasonal models look three to seven months ahead. They are most skilful in the tropics, where ocean temperatures drive the weather. Over the UK, whose weather is steered by the fickle path of the jet stream, the skill is modest. A seasonal forecast of a milder winter shifts the odds a little; it does not rule out a fortnight of snow.

Part of the warm leaning seen in many seasonal forecasts is the climate warming: the model's reference years include cooler decades. That is worth knowing when you read one.

How to spot a trustworthy long range forecast

Be wary of any forecast that gives a specific temperature for a specific day a month away, or confident headlines about a whole season's snow. Trustworthy long range forecasts show their uncertainty, compare with what is normal, explain where the numbers come from, and change their mind when the models do. That is the standard we hold ourselves to.