Capitalism for Cutie

One of the 20 best sentences I've ever read in an academic economics journal:

E/P − Y (the Fed model) has some power, but again only because it is a poor man's E/P.

—Clifford Asness, "Fight the Fed Model: The Relationship between Future Returns and Stock and Bond Market Yields," Journal of Portfolio Management 30.1 (2003): 11–24 [I don't have the exact page number].

While I'm on the topic: I honestly don't understand how economists and the like can get a computer to draw a line of best fit through some points (i.e. ordinary least-squares regression), compare the values on the line to the observed values, and then, if the two sets of values are close to each other, claim "forecasting" power. They didn't forecast anything yet — the only test of that will come, you know, tomorrow and the day after and beyond. "Forecasting" the past is way easier than forecasting the future, especially when you get to take into account all the historical data, not just the stuff that would have been available to you at any given point in the past. Which suggests a much more interesting test, one that I'm sure has been thought up (and discarded?) by many others but which I nonetheless don't remember encountering in the past. To wit: you have some theory about how, say, the P/E ratio predicts stock returns. What would that theory have predicted the market return would be in 1972 — given the historical data available in 1972, and nothing else? How much better or worse does the prediction become when you get to do the regression based on the whole span of historical data? Basically, the idea would be to forecast historical values using only the information that people in the past would have had on hand. Then you'd know if, in the past, you would have been a good forecaster, which might give a better sense of whether, in the future, your predictions will be worth a damn.

Is this a standard technique? an erroneous one? What am I missing? Is it the fact that, throughout the early to mid-20th century, most stock traders had access to rudimentary time machines?

Google Image result #33 for "time machine":

LIVEBLOGUPDATE: Yeah so when I read more of the quoted-from article, I saw a reference to "out-of-sample forecasting." After googling this for about a minute and a half, I'm fairly certain that it's exactly the kind of thing I was proposing directly above. Some technical details can be found in a Word document that came up as the second search result. I wish I could take this as a sign of my brilliant econometric intuitions, but, like I said, the basic idea is pretty basic. There goes my fake Nobel prize (all the econ ones are fake — you knew that, right?).