# How do 14-day forecasts work?

42 views
0

I can see trends over the next day or two being predictable based on current weather… but how is it done for 14 days out?

In: 14

These days, really fast computers.

The Americans have the Global Forecast System (GFS). The Europeans have the European Center for Medium-Range Weather Forecasts (ECMWF).

Other countries run less powerful systems, and other systems exist that are used only in the next 1-3 days.

All of them run a *ton* of physics and math calculations. Experience has taught us that if, say, a super typhoon were to move into the North Pacific near Japan, five days later the weather pattern will change in a relatively predictable way in the U.S.

From this you can at least get a broad idea of whether or not it’ll be warm in a certain region, and maybe even if it will be rainy. This idea is called [teleconnections](https://www.cpc.ncep.noaa.gov/data/teledoc/teleintro.shtml).

Out to 14 days, you’re relying more on past history and hoping the math gets you close. But weather is unpredictable, and little errors in the Day 2 forecast spiral into huge errors by Day 5.

Meteorologists try to work around this by running a whole bunch of 14-day forecasts, then using the average, and/or discarding a few because the forecast is unreasonable (in their opinion).

This gets you… closer, but the nature of the problem means it will still be significantly wrong a lot of the time.

In the same way. Atmosphere is divided into cells with various variables and supercomputer then calculate the relations and effects between the cells. Though it is very much just an educated guess for the longer periods of time.

Weather models are quite precise and good these days, but in the end they are still just models, an approximations and representations of reality. Basically only forecast you should look at is for the 3 days, maybe for a week at maximum.

The short answer is, sometimes they just don’t. It’s a predictive model. It can be as simple as looking at previous weathers, trying to find patterns in it, and trying to build a model that predicts weather in 14 days based on what you observe (in terms of atmospheric pressure, wind, temperature etc) during the current day.

The model building part isn’t made by having a human scratching his head trying to find patterns in weather, of course. It can mean just slapping a neural network on it and having it try to predict weather from all previous weather data it read through.

These models might be more or less accurate, more or less capable to accurately predict what’s gonna happen in 14 days. If they usually fall short of an accurate prediction, feel free to take these forecasts with a grain of salt.

The weather is a chaotic system, which means that unavoidable small errors we make in setting up our forecasting computer models — for instance not having observations at every single point on the entire globe — will get worse and worse over time, until our simulation is totally different than the real world We can minimize these errors by using very good models and observations, but that just pushes the problem down the road. You can’t beat exponential growth.

These graphs show some estimates of forecasting skill from modern weather simulations as a function of how many days into the future we’re trying to forecast. Two, three, four days into the future they’re great, but they get worse real quick. The horizontal line on the graphs is a rough estimate of when the computer simulation is no better than a random guess.

[https://www.researchgate.net/publication/321808794/figure/fig3/AS:[email protected]/HSS-values-based-on-ECMWF-reforecast-predictions-of-anomalous-AR-activity-as-a-function_W640.jpg](https://www.researchgate.net/publication/321808794/figure/fig3/AS:[email protected]/HSS-values-based-on-ECMWF-reforecast-predictions-of-anomalous-AR-activity-as-a-function_W640.jpg) [https://www.ecmwf.int/sites/default/files/AR2016-Delivering-regime-transition.jpg](https://www.ecmwf.int/sites/default/files/AR2016-Delivering-regime-transition.jpg)

So you can see that 14 days is about the limit. The forecast is right a bit more often than it’s wrong, so some people want to see it, but you definitely can’t count on it being at all accurate.

It’s also interesting how this has changed over time, as our weather observing network and computer simulations have improved. This graph shows the change in forecasting skill year after year:

[https://img.apmcdn.org/75e700b8a14453156922534a92480e00f7735f94/uncropped/3b27a9-20200102-european-model-ecmwf-forecast-accuracy-since-1980.jpg](https://img.apmcdn.org/75e700b8a14453156922534a92480e00f7735f94/uncropped/3b27a9-20200102-european-model-ecmwf-forecast-accuracy-since-1980.jpg)

You can see that despite huge improvements in technology, the gains are noticeable but not huge. Today’s forecasts 10 days into the future are about as good as the 7-day forecasts we were making in the 1980s, and our 5-day forecasts are a bit better than the 3-day forecasts from the 1980s.

Oh, and also, these graphs just show the predictability of the large-scale wind and pressure patterns. Predicting whether it’s going to rain at a specific spot at a specific time — which is what humans mostly care about — is even harder.

For more on chaos and how it affects future predictability, check this out: [https://www.aps.org/publications/apsnews/200301/history.cfm](https://www.aps.org/publications/apsnews/200301/history.cfm)