On the contrary, it is you who are unable to grasp the significance of correlation regardless of causality.

If, in a large population, certain attributes are correalted with negative outcomes, then if I select a sub-population that excludes population members with those attributes, then axiomatically I have reduced the proprtion of negative outcomes. The only uncertainty is whether or not correlations will persist into the future.

It doesn't matter whether those attributes are causal or coincidental - whether it is a combination of conventional credit scoring factors or the number of letters in last name combined with the day of the week on which the birthday will fall in 2013.

The trap that your thinking falls into (either conciously or sub-consciously) is being overly concerned with which members of a sub-population would in reality default if no action were taken. From a risk anaysis perspective that is unknowable and irrelevant. A perfectly valid way to reduce the risk of negative outcomes within a population is to identify subpopulations the collectively exhibit higher than average negative outcomes, and to exclude that sub-population.

Beacuse the past is not a certain predictor of the future, it is not possible to be certain that excluding such a sub-population because it's past attributes were correlated with past negative outcomes will reduce the proportion of future negative outcomes, it is highly probable - provided that the past correlation has been reasonable constant over time.