A1174
Title: New modelling method based on data augmentation applied to binary variables
Authors: Aneta Ptak-Chmielewska - SGH GV Warsaw School of Economics (Poland) [presenting]
Pawel Kopciuszewski - ING Hubs Poland (Poland)
Abstract: A new modelling method based on data augmentation applied to binary variables is presented. The proposed approach extends the three-phase modelling method applied to continuous variables. The effectiveness and properties of the method are demonstrated using credit risk data from a portfolio of mortgage loans and synthetically generated data. The method is based on data augmentation and construction of a model in three phases. In the first phase, a model is built that predicts the difference in the target variable between two observations. The exogenous variable for the phase 1 model is defined as the difference between target variables for pairs of original observations in the raw dataset, and an appropriate regression model is constructed to predict the differences between observations. This technique increases the number of pairwise observations between the original observations and focuses on developing a pattern of pairwise observations rather than searching for a profile of the original observations. The second phase involves the construction of a residual model, and the third phase involves the selection of a set of comparable observations to predict the target variable. The proposed three-phase modelling approach can significantly improve the quality of the GLM model.
