Let me present two really good illustrations for why not to trust models more than a week or so into the future, and especially beyond 10 days.
The Global Forecast System model’s preposterous flip-flop
On Monday, the GFS model forecast 15 days into the future (for Dec. 19) predicted Ice-Age-like subzero temperatures (in the afternoon) from the Carolinas to central Georgia and a raging snowstorm. Such a scenario would be unprecedented in historical weather records and is not at all realistic.
“Comedy gold: GFS 18z transitioning from our current climate to Snowball Earth,” tweeted Ryan Maue, a meteorologist at WeatherBell Analytics. “Not sure this is even possible.”
The frigid forecast is all the more absurd when you consider the forecast made for the same point in time the day before: highs in the 60s.
In other words, the model’s forecast temperatures shifted 70 degrees in a single day.
The pachinko analogy
The National Weather Service forecast office in Kansas City, Mo., developed a really helpful analogy that shows how and why a forecast can change so radically.
Consider a game of pachinko, in which a pinball has a starting position but can go in any number of directions until it reaches some destination of a given point value. The pinball is like a weather event that can also take many twists and turns, over a period of many days, before you know where it is headed and how severe it is going to be.
If you want to assess the likelihood of a certain outcome in pachinko, you could take a pinball and drop it in the same place multiple times and then see how often it finishes in the same spot. But in most cases, you would find the ball landing in many different locations — indicating low confidence in any particular outcome.
Meteorologists conduct this same sort of exercise using what are known as ensemble members, which are alternative versions of a given forecast simulation with some small tweaks. But when running such modeling experiments at time frames beyond five to seven days, forecasters usually see quite a bit of spread in the actual outcomes — which is why confidence in such forecasts are usually not all that high so far into the future.