베이지안 머신 러닝을 이용한 은행권 주택담보대출 예측

This paper provides a mortgage loan prediction method and density forecasts. To this end, we develop a Bayesian machine learning algorithm based on conditional Bayesian model averaging. In this approach, the uncertainties of the parameters, models, and prediction variables are incorporated in the predictive density simulation. The first learning stage is the variable selection. Among a number of potential prediction variables we choose the prediction variables which improve the one-month-ahead predictive density accuracy compared to an uni-variate model. The second learning stage is the model choice. We evaluate both direct and indirect prediction models using the selected prediction variables. The model weights are obtained by the relative out-of-sample density prediction performance, In the last stage, the predictive density forecasts are constructed using the model-specific density forecasts and model weights. According to our empirical results based on the monthly data ranging from December 2007 to October 2016, the variable and model uncertainties are substantial, and the relative importance of the variables and models are dramatically time-varying. It is also found that recent rises of the mortgage loan are attributed to the changes in the apartment price index. Finally, the mortgage loan growth forecasts are decreasing in the forecast horizon.