# Loan level model to understand drivers of mortgage prepayments

I am following up from my question here. As described there, I'm trying to assess the drivers of CPRs for a type of MBS. However, I want to understand, how a loan-level model of such a relationship would look? Specifically, how does one model the drivers of prepayment at the loan level, as a typical regression will not work given that the data is both longitudinal and cross-sectional in nature. So, in this case, is there a tried and tested regression or other procedure that is used? I am leaning towards a Fixed Effect model.

• Big investment banks pay millions of dollars to develop prepayment models. Bloomberg has several dedicated staff who do MBS modeling full time. No one is going to give a good model for free, and any free model will be garbage
– user50421
Oct 16, 2020 at 23:21

Probably the most established (thus far...) academic approach to this problem is to use a multinomial logistic regression (the search terms "prepayment model multinomial logit" should turn up dozens of papers). Longitudinal effects are captured in some of these models by including time-dependent explanatory variables such as an index that captures the availability of mortgage credit, the vintage of a loan etc.

On the other hand, it is not clear that prepayment data satisfies a crucial assumption made by the model, namely the "independence of irrelevant alternatives." Various extensions have been proposed to the baseline MNL framework to handle this but these introduce additional layers of complexity to the estimation process.

Unfortunately, there is no canonical reference (that I know of) that provides a guided tour through these nuances in the context of prepayment modeling. Going through the modeling details of some of the papers found by searching and working through their references and also asking specific questions appears to be the only way right now.

• Thank you for the answer. I have the actual CPR data and my analysis is more on the explanatory side, assessing driver of this actual CPR, rather than predictive. I'm aware of multinomial logistic regressions being used for predictions of prepayments, within a Markov Chain framework. However, in this case, Im just trying to understand the drivers of prepayments and that's why I'm more concerned about analysis of the coefficients of the regression, etc. I want to use a loan level model so that I can capture as much heterogeneity in the information as I can.
– Jojo
Oct 11, 2020 at 21:08
• And as this is a loan level model, the data is longitudinal and cross-sectional in nature, given I have data for more than 1 loan per timestamp. In such a scenario, I was under the impression there is a different set of regressions associated with panel data, like the Fixed Effect Model?
– Jojo
Oct 11, 2020 at 21:08
• Just to add: My Y-variable here is the actual CPR, rather than probability of transition.
– Jojo
Oct 11, 2020 at 21:15
• I'm not familiar with using a Fixed Effect model in this context but that may just be my ignorance. Prepayment modeling is a very large and complex area with several different approaches. Oct 11, 2020 at 22:11
• This thread is becoming too long so last comment. There's no statistical approach that is going to be able to deal with the full richness/complexity of prepayment data over time. If you have an idea, try it, explore its shortcomings, and then try to see how you can patch them. Oct 12, 2020 at 10:21