Research on debt recovery My final year project is on debt recovery data for a debt collection firm. Data such as original/current balances,payments made,DOB, number of contacts made,whether or not a debtor has made insuarance payments are given 
We need to model the given problem in order to predict the collectability of debts of a given debtor. Only approach I started on was a logit model. Any other ideas or hint on possible ways to approach the problem would be appreciated.
 A: Another method is SVM (Support vector Machines). The classifier function constructed during the training completely decides the differences in data between the classes. We can do this using SAS/R/Python.
If you have two/more categories in output variable I suggest you to look at Discriminant Analysis as well. Hope this helps !
A: Firstly, I would suggest looking at what work has been previously done on similar problems. For example, a bank offering a mortgage looks at similar sorts of data and then predicts how likely it is the person will be able to pay it back. This should give you ideas on how to tackle the problem.
Logistic regression certainly seems like a good idea. Another thought is maybe looking at some of the multivariate data analysis methods to classify people into different groups, based on the data you have. Specifically correspondence analysis and its extension, multiple correspondence analysis, are the counterparts of principal component analysis for categorical variables, and may be of interest.
A: A logistic regression is a good place to start. However, I would also recommend you look at modeling in the area of loss given default (LGD), since LGD = 1 - recovery rate (RR). You sound like you're trying to model a recovery rate, so a simple transformation of LGD may be one way to attack this.
In LGD modeling, it's commonly observed that loss distributions are bi-modal; you tend to lose very little or a lot. So, think about that before getting too far into using a logit. I don't know what kind of data you have/don't have, but if the data doesn't fit a logistic model, be prepared to take another approach. No use in trying to fit data into a poorly spec'd model.
