# Why isn't prepayment modelling done on actual prepaid amounts?

When modelling prepayments in Securitized products, why is it that the standard model involves a transition matrix (Markov Chain) framework, where the probabilities of transitioning between different states of delinquency are modelled? I have searched and have been unable to find any literature corresponding to models that are used to directly predict the actual prepayment amounts ($amount) for each time period, given a dataset containing that information. As this information is directly available why isn't a regression fit directly to that data to predict the actual prepayment amount? I would like to understand why this seemingly easier method has been eschewed in favor of the Discrete-time Markov Chain framework for prepayment modelling (which requires logistic regressions for each transition that is modelled). If there is literature on such modelling, would be grateful if someone could provide the corresponding links. • Usually, there (at least) two approaches: one is a rational decision model, effectively stochastic optimization / swaption pricing (simplest case); the other is behavioral modeling where you do precisely what you suggest: Analyse historical prepayment data sets and try to estimate behavior. – Kermittfrog Nov 15 '20 at 19:14 • @Kermittfrog Thank you. Is there any literature you could point to that lays out the behavioral modeling or typical procedures that you could speak to that are used for this? To do this, I am looking into using a two part model: 1) Which models the probability that there will be a prepayment (through a logistic regression) and 2) which models the actual prepayment amount conditioning on there being a prepayment. I do this because, when looking at loan level data, for the dependent being the prepay amount (in$), most values are 0 and the loan level data is in a short panel format. – Jojo Nov 15 '20 at 20:13
• @Kermittfrog But I would like to see how others have modelled this and what explanatories they have used, etc. – Jojo Nov 15 '20 at 20:15
• unfortunately, I have only seen some models applied within banks. I think I am not allowed to talk about those, but it is fair to say that it’s often not the modeling that impeded our results but the data quality. So we’ve applied quite simplistic models most of the time. – Kermittfrog Nov 16 '20 at 7:14