http://www.federalreserve.gov/boarddocs/Speeches/2002/200211...
The reason is that given any dataset, you can come up with thousands of stories that explain the data perfectly, but have no predictive power. This is also called 'overfitting' in machine learning contexts.
However, a prediction made before events can use those events as evidence of the correctness of the theory.
What is the evidence for the theory in this article? Well, it is that it explains past events. But those events can't be used as evidence for the theory, and if that is all there is, this is a 'just so story'.
Another way to say this is: "The fact that a theory does not conflict with existing observations is not evidence that it is correct."
Edit: If you mean that you need to state the theory before you make the observation for the observed data to be valid evidence for or against the theory, you are correct, this is very nearly contrapositive of the scientific method.
Two things immediately come up when looking at economic theories:
1) a LOT of them (the ones "followed") are "by" high-level Fed officials. Biggest example: Keynesianism.
2) They all seem to have the same conclusions (government needs to spend more, either directly or through the fed or finance dept)
3) because these people want to but can't be bothered to manage an economy (meaning hundreds of thousands of small companies) they massively advantage large corps over everyone else.
That's complete nonsense. Theories are almost always made after collecting observations.
They may be confirmed using future observations, or by statistical analysis, however.
There is nothing wrong with looking back to see if a variable explains an observation. However, the true test of the hypothesis is predicting future events.
Because we can't trust people, future is a safe and sufficient choice. It's not necessary. (That's why machine learning people use hold-out sets.)