Custom Prepayment and Default Modeling

Projects:

Client: Commercial Bank - Mortgage Group
Project: Development of a Proprietary Prepayment Model
We designed a prepayment model for a New York commercial bank's whole loan mortgage portfolio. To make full use of loan-level data, we worked extensively in the area of survival/duration analysis; especially hazard models using time-varying covariates. We then fitted parametric distributions, including the Exponential and Weibull to a data set with over 500,000 time-varying records and investigated the significance of the various covariates. Our work involved accelerated and heterogeneous failure time models, censoring and truncation, proportional hazard models and maximum likelihood techniques to estimate a prepayment function. In the end, we proposed ways to maintain and upgrade the model, including the possible use of mixture distributions.

Client: Large South American Bank
Project: Build mortgage prepayment model based on the country's mortgage market
The client wished to improve the pricing of their mortgage securities in the US market and thus needed to better understand prepayments. In addition, they intended to form and split off a mortgage insurance (default insurance) company. We built a custom prepayment model that was to become an essential part of the asset-liability management system of the (future) insurance unit as well as for future securitization/valuation.

The project entailed the normal tasks of data analysis, such as discovering relationships between the variables. We also visited with builders, construction people, housing unit managers, mortgage loan officers and insurance risk managers. We researched transaction costs and risks not found in the US market and began to understand psychological differences in borrower behavior.

Client: Large Mortgage Bank
Project: Build a model to forecast prepayments on a loan level basis
To develop the 30 year fixed rate prepayment model, we first performed a comprehensive data analysis study on the client's 1.5 million loans. The study identified potential independent variables and indicated non-linear or interactive effects that might impact subsequent modeling steps. Next, we developed a preliminary model based upon measures of statistical fit as well as through analyses of various cohorts of loans. Subsequently, we introduced additional variables into the final model after evaluating new functional forms and interactions.