It's essentially scenario forecasting. It's kind of like if you're planning to take your family to the beach this weekend for a nice holiday of fun in the sand. Your initial conditions are that the car works, your boss gives you the weekend off, the forecast is warm and sunny, and you are feeling good.
If you're like Billy Bob, you don't give it a second thought. But if you're like Geoffrey Von Enrich, III, you think it through. What will my weekend look like if my boss makes me work? Maybe we'll stay in town. So we'll maybe we'll end up eating a nice evening meal and going to a movie. Well, what if the movie is sold out? Maybe we'll buy tickets to a concert. Or, maybe I'll send my family ahead to the beach. Ok. Well, what if it rains? Maybe we'll go to the beach anyway but go to the indoor water park. What if my car breaks down? Maybe I rent a car and go to the beach. Or maybe I stay at home and watch movies and play games instead. What if I'm sick. Maybe I stay in bed or send my family ahead to the beach. Or maybe they stay home with me.
How many scenarios end up at the beach? What's the spread on the other outcomes? There's value in gaming it out. If most scenarios still end up with people going to the beach, then you have a pretty solid idea of how the weekend plays out. But if a couple of key things are changed that cause you to end up in a different place, then there's a pretty solid risk to the beach trip.
Meanwhile Billy Bob has no idea what to do when his car breaks down and his boss asks him to come in. He walks to work in the rain and gets the flu.
Anyway, this might be a terrible example. But the point is, models are imperfect and are prone to large errors out in time. Tweaking the initial state of the atmosphere and observing whether or not those tweaks cause large changes in the outcome downstream, provides a good insight as to whether or not you should have confidence in the general pattern being depicted out in time.