I'm not sure what you mean by "incorporate interest rate". I did that by making a simple logistic regression model to predict default rates, using all the data for completed 36 month loans through 2013. I included a few parameters that seemed to contribute substantially to the variance; one of the best predictors was "inquiries in the last 6 months". I also included "interest rate" and a few others. So then I had a regression equation that let me determine the probability that any given loan would default. Of course it works well for the historical data (because that's what the model is based on). If LC recently bumped up the interest rate for borrowers with a high number of inquiries, the model might still work roughly OK because it includes interest rate as one of the predictors. (Although using "category", i.e., A1-G5, as a predictor actually works better than interest rate, presumably because it's not subject to the variability in interest rates that LC introduces by responding to market pressure.)

HOWEVER, this does not take into account the possibility that LC actually changed their underwriting for loans with higher inquiries. Suppose they rejected more of them unless the borrower met other requirements, such as having a higher income, lower DTI, or some other as yet non-transparent factor. Then the influence of inquiries on default rate would be fundamentally different today than it was in the last few years (because the population sampled by LC is now different: they are accepting and rejecting a different group of borrowers with different characteristics, among which "inquiries" is no longer strongly related to default rate), and the model would overpredict default rates. I'd be hurting myself by using "inquiries" as a criterion for rejecting notes.

So, while I completely agree with you that historical data is the best way to predict the future, those predictions assume some consistency between the past and the present. Because LC periodically changes its underwriting and rate-setting alrgorithms, there is less consistency - perhaps a lot less. So predictions are less valid, meaning that the risk of using them is greater.