• big data
• neural networks
• machine learning
• forecasting
• macroeconomics

This project tackles the issue of dimension reduction and forecasting with high dimensional macroeconomic data. The most common dimension reduction technique is Principal Component Analysis, which forms the backbone of Dynamic Factor Models and is based on linear transformations of the data. Yet, scholars have argued that non-linear structures in the data could be exploited to obtain an improved representation of the original data and more accurate forecasts. In this project, I experiment the use of neural networks, part of the arsenal of techniques coming from deep learning, as a dimension reduction method that exploits non-linear interactions. Using monthly data from the United States from 1960 to 2019, I estimate factors by Principal Component Analysis within a Dynamic Factor Model and also by training two autoencoders with differing depth. Subsequently, I generate factor-augmented forecasts of US Industrial Production and compare the results in order to assess both methodologies. A basic one-layer autoencoder is shown to yield factors that are able to collect a significant amount of features in the data, comparable to a Dynamic Factor Model, and which improve the forecasting accuracy with respect to the Dynamic Factor Model at all horizons, especially at the six and twelve steps ahead. On the other hand, a deep autoencoder with five hidden layers struggles to find an improved lower-dimensional representation, but the factors obtained may still lead to noteworthy gains in forecasting accuracy. Further research may be conducted in order to improve the performance of these networks, reducing overall loss and overfitting and increasing the amount of information retained by the factors, while further experiments could see other architectures being exploited to obtain more accurate forecasts.