Forecasting Mortgage Demand: An Application of Traditional Methods, Machine Learning, and Neural Networks

Abstract

Demand forecasting relies heavily on traditional methods with well known limitations. Improved accuracy in predicting demand for mortgages, whether for purposes of purchase or refinance, is critical to profitability in home lending. To overcome obstacles to prediction using nonlinear relationships between variables and to long-term accuracy, we apply time-invariant machine-learning methods such as random forests. We also perform time-series analysis with a wide variety of deep learning architectures, including convolutional and recurrent neural networks. Time-series analysis through deep learning produces the most accurate results. Even shortcomings in forecast accuracy can reveal tacit changes in relationships among household-level and macroeconomic predictors of mortgage demand.

Keywords: Demand Forecasting, Deep Learning, Mortgage Demand, Purchase Demand, Refinance Demand, LSTM, RNN, Production Capacity, Uncertain Demand, Decision Science