Deep Learning architectures in macroeconomics and finance

My research focusses on nowcasting and forecasting macroeconomic variables such as GDP, inflation or rate of unemployment in the United Kingdom and other developed economies using deep neural networks (DNN). Hereby, this thesis makes three key contributions to current research: First, in the context of macroeconomic forecasting, the application of deep neural networks is not fully explored yet. Secondly, a data-driven algorithm is presented to derive an optimal network structure for a multilayer radial basis function network. Thirdly, to the best of our knowledge, no research has yet been conducted in the field of macroeconomic and financial forecasting using datasets with up to 50,000 variables. This thesis therefore presents an unprecedented empirical study which also includes state-of-the-art innovations in methodology. It aims to not only impact the academic discourse but also to make relevant contributions to the non-academic world.
Macroeconomic and financial forecasting plays an essential role in central banking, policy making, regulation and in the free economy. It is omnipresent in financial and policy news since its outcomes have wide effects on society overall as well as market participants or private companies more specifically. What is the United Kingdom's GDP expected to be next quarter or is the Bank of England on track to hit its target for inflation? One of the most important aspects of those questions is that macroeconomic statistics are usually only published with a lag which means that the information regarding a certain quarter or month only becomes available delayed opposed to being accessible "live". However, in a globalised and fast-paced environment, having accurate now- and multiple steps ahead forecasts available timely is extremely valuable.
After the most recent financial crisis in 2008, it became obvious that current forecasting models do not pick up all relevant information to recognise potentially catastrophic changes that are about to happen in the economy. The need to be able to forecast the near future more accurately is therefore emphasised. This research combines large amounts of data including (macro-) economic and financial variables as well as "Big Data" such as Twitter sentiment or Google Trends data. To the best of our knowledge, such extensive datasets have not yet been applied in the macroeconomic forecasting literature partly due to the problem of overfitting.
In terms of methodology, this thesis focusses on Radial Basis Functions (RBFs) as activation functions in the hidden layers of the deep neural network. The reason for focussing on RBFs is the advantageous possibility of choosing suitable parameters for the hidden units without having to perform a full nonlinear optimisation of the network. While RBFs themselves are not new in literature, this thesis makes an innovative contribution but constructing deep neural networks that have RBFs in more than one layer. The advantages of deep neural networks over conventional factor or time series models include their ability to approximate any nonlinear function, incorporating a large number of variables as well as being non-parametric, meaning that they do not require any specific underlying economic theory.
First results are expected to be available in early 2019.