Your observation is correct Al, machine learning models are trained to predict the past, and there is an assumption that success in predicting the past correlates with success in predicting the future.
This assumption gets routinely violated in real life, and correlation between success in predicting the past vs future is not 100%. But it is not 0% either. And it is way better than trying to predict the future based on your intuitive understanding and trying to code them up as rules. So there are two definitions of a “good” model. One is that the model is close to making 100% accurate predictions, the other is that it is better than its alternatives. And I would say the second definition, while not completely satisfying, is good enough to keep machine learning in business.