I have some knowledge on survival analysis, but limit myself to healthcare background. I think, it is not totally the same as your topics on mortgage payment. I will say I don't have a solution on your question (even I am very interest in this field now) but just to share my comments.
Doing survival analysis, we usually assume that the subject will die (experiencing event) eventually. It is proper for one’s life but not for mortgage default problem, because many people won’t have such problem at all to then end of payment, that may serve the reason as why, as you mentioned, a lot of censored in your data. If we meet such problem, we usually think that we may not have a long enough time window to observe the subject. Too high proportion of censored data indeed cause problem on survival analysis. I usually compare that with a very low number of events in logistic model.
I do not fully understand what is a noisy estimated survival curve? You mean, the predicted confidence intervals are too wide for the curve?
I will read more on this topic since it may be related to my work now.
jingju

Thanks for your reply.. Your understanding are definitely right about the cause of small number of default events. The censor rate is above 90% in my sample data, so I just pull all the default loan into the sample data now, the censor rate will drop dramatically.. Then I will use some calibration method to get the score ..
What I mean about the noisy is my curve seems have large variance. I apply kernel smoothing, the curve seems better now..

Thanks for your reply.. Your understanding are definitely right about the cause of small number of default events. The censor rate is above 90% in my sample data, so I just pull all the default loan into the sample data now, the censor rate will drop dramatically.. Then I will use some calibration method to get the score ..
What I mean about the noisy is my curve seems have large variance. I apply kernel smoothing, the curve seems better now..

good..