An approach to securitisation in Europe NPLs- machine learning model field lab project Nova SBE | moody’s analytics

As a consulting project, we were proposed to develop a neural network (NN) to predict mortgage states in one year, based on the paper ‘Deep Learning for Mortgage Risk’ by Justin A. Sirignano, Apaar Sadhwani, Kay Giesecke (2018). We developed a neural network model with the aim of being able to capture the relationships between the different variables, with respect to each other and to the response variable (the loan status in 12 months), better than traditional classification methods, such as logistic regressions, which constitute the benchmark set. Data was provided by Moody’s, relating borrower, property and loan/financing characteristics for several mortgages over several periods in time (over 350 thousand mortgages). The purpose of our model is to predict the probabilities to transition to different states at a certain point in time. The best results were obtained with a 10 layer, 500 nodes per layer network. The model can identify a large portion of defaults. At the cost, however, of a general overestimation of the default rate over the years. The capability of identifying loans that will be in arrears is also acceptable, with, again, an overestimation of the verified rate. Variables relating to borrower characteristics and history as well as financing are found to be the most significant