In a "model" for prediction, the model can be "made" at any time(not necessarily prior to the events it is used to predict).
In this case with CU, there are certain parameters they use to say if X, Y and Z are present and A, B and C are not present, then the outcome of the race favors the incumbent. If X, Y and Z are not present and A, B and C are present then the election favors the opposition to the incumbent.
They then can "Look Back" with their model and apply these parameters to each election and "ask" the question whether the model was accurate in predicting the outcome of that given election when their "predictors" were applied.
Having done that, they can then say their model was accurate or not accurate for a given historical election. This is what they meant by.. "not wrong since 1980" , even though their model was created many years later.
As an addendum...just because the "model" was 100% accurate since 1980, doesn't mean it will be correct for the next 8 elections. This is all based on odds or game theory and it lays the odds heavily in their favor with the 8 consecutive correct predictive history.
This is the same "modeling" myself and others do to look back at the historical record on EX runs. Even though the "model" is correct 80-90% of the time, and in some cases 100% of the time over the last 8 Q runs. I have lost early this year on EVEP as it failed to run in May(the only Q in the last 15Q's it hasn't run).
So I got frustrated because I lost money in May and did not play it during these last two Q's when it ran up 4 dollars to the Aug EX and 3 dollars to this last Friday for its EX tomorrow(Go figure). You have to stay in the game with your model, otherwise you skew results helter/skelter when you have an efficient prediction model.
If someone would take my bet I would bet heavily for CU's candidate using their model(if you'd give me 1/1 odds) ,with its history. Problem is, the bookies would lay down steep odds against my bet if they were inclined to believe CU's model. 1/8 odds means my $10,000 dollar bet would give me $1,250 profit, for a win...and a $10,000 loss if I lose..I don't know if I trust the CU model that much....;-)
Hey Doc. I get all that (probably even more clearly now than before your great explanation), but, in a way, I think it helps make my point. Isn't it similar, as you said, to what you are doing with divy runs in that you look at historic conditions, find a set, or two, that happen in a quarter when a good divy run had occurred and look for these conditions in other quarters, too. If they are found many (most) times, we would start to think there was a correlation between the conditions and a divy run.
If I teach my model that when x is true, y results, and then input data x from a prior election, of course its gonna return y, because that's what I told it to do. Testing this model against the same data points that were used to construct it only tests that the model is built the way I wanted it to be built, not that the data is actually causal to the result.
Admittedly, the more historical events that match the thesis and the fewer that don't, the more likely one is to "expect" an accurate prediction for some future event. The CU model has been built on 8 presidential election results, three of those had no incumbent, and, I think, only two instances of an incumbent loosing. That seems like pretty small sample to be risking $10k on, no matter the odds.
Interesting topic, but I wonder how much each person's passion for tomorrow's outcome influences their willingness to latch onto this prediction. Such is the case in almost all influence peddling...when you are "preaching to the choir", you don't really have to make sense.