Macroeconomic and Financial Market Analyses and Predictions through Deep Learning

Since Hinton, Osindero, and Teh (2006) developed the fast learning algorithm, deep learning has been a set of powerful tools that has recently achieved impressive performance across a wide spectrum of industries as well as in academia. For the macroeconomic and financial variables, however, more elaborate approaches need to be taken due to the unique latent features of them. In this regards, we propose novel approaches to apply deep learning to the predictions of time series variables in those fields. Specifically, we suggest ensembles of neural networks and Bayesian learning to estimate the posterior distributions of the forecasting outcomes as the out-of-sample forecasts. Examples are provided with predictions of monthly custom clearance exports from Korea and daily Korean won-US dollar exchange rates. The prediction results show that deep learning approaches are prevail even with non-granular data sets which normally used for the conventional econometric models.