To: Bonaparte
"He's applied his model all the way back to 1916 and it holds up surprisingly well."
What he is doing is called "data mining" or "retrofitting." If I have enough time and resources, I can create a formula that is consistent with any historical sequence (stock prices, elections, wars, etc.).
Unfortunately, the next case may not be consistent with the past, so I may need to adjust my model again. The ONLY way to know if such a model is of any real predictive value is to see how well it performs prospectively (not merely how well it fits past data).
He has used his formula prospectively for the past few elections, but "back-testing" to 1916 doesn't show anything other than that his formula is consistent with the period he used to develop the formula. It ought to work for that period.
30 posted on
08/14/2004 4:45:58 PM PDT by
labard1
To: labard1
I understand about retrospective studies but it seems that if pervasive patterns are found, they should at least be acknowledged. Nobody suggests that he's got a crystal ball.
To: labard1
How do you "know" that he is "data mining?" It is standard procedure to make one's predictions "out-of-sample" (what you call "prospective.") One simply fits the econometric model without the event to be forecast, e.g. fit the model with all past elections except one. Fair may well have done this to "predict" each election since 1916. Or he may have not. Unless you know the facts, you should not be dismissing his research.
FreeRepublic.com is powered by software copyright 2000-2008 John Robinson