Accounting for heterogeneity in aggregated mortgage pools with the aim of forecasting prepayment / default risk

Abstract: I investigate the use of segmentation of aggregated mortgage pool data in a Markovian setting. I aim to increase out-of-sample forecast accuracy of mortgage performance. To this end I implement a Bayesian multinomial logit model with an adaptive Metropolis-Hastings algorithm to sample the posterior distribution. I devise a decision rule to map the multinomial probabilities to mortgage performance forecasts. I find that segmentation based on mortgage characteristics not directly defined in the data set, but that can be properly formulated, increases model fit and forecast accuracy for non-performing loans. I also discover that performing loans demonstrate a remarkable level of homogeneousness.