They have lots of models that are built on how things worked in the past. We have a hundred years of detailed weather data, and we can (and do) use that data to build models that are predictive withing certain ranges. We can test them by (for example) picking a date in the past in a given location, input the recorded weather conditions in the model and see how close it came to predicting what actually happened (we check it against the recorded data). Do this enough, and we learn how to tweak the models to increase their ability to predict conditions. And models do get better as a result.
That said, even the best models struggle to predict anything reliable more than a few days out (other than “it’s gonna be hot in the summer” or “it’s gonna be cold in the winter”). There’s just too many variables and too much uncertainty more than 72 hours down the road. Weather modeling is a classic case of applied chaos theory because huge differences in “outcomes” (what the weather actually does a week from now) happen with very small differences in initial conditions (what’s going on right now).
Latest Answers