Say you have two negatively correlated variables (GDP and UR) in a multiple regression to predict bond defaults. Is there any reason you wouldn’t want to use negatively correlated variables in the same equation? Is there a recommended cut-off for the Pearson coeffecient?
If there is multicolinnearity, then you might find significant relationships from ordinary least squares when they do not actually exist (standard errors are larger than OLS would show). If the correlation isn’t particularly strong, then it’s not a very big deal. You cannot use OLS when the correlations are strong though. More info here: http://en.wikipedia.org/wiki/Multicollinearity There are bigger problems when predicting defaults, though. Namely, you should be using some sort of probit or logit model, rather than traditional regressions. Eric Falkenstein also has an interesting paper on forecasting corporate defaults: http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1103404