About
This blog was created and used to track the progress of my MEng final year project. Currently it contains all deliverable for the project, including the report, presentation from the thesis defence and experimental results.
Abstract
The field of macroeconomic forecasting has been relying on classical methods, with limited expressive power. In this project a new method for macroeconomic forecasting based on deep neural networks is proposed. It leverages the innovations in the field of deep learning to produce accurate and efficient forecasting models, and attempt to answer a number of research questions: Can deep neural networks performance surpass that of classical econometric models? How does preprocessing affect the accuracy of forecasts? Can machine learning algorithms automatically tune the parameters of deep learning models?
In order to answer these question, a robust experimentation framework is implemented, which through search optimization algorithms, performs automatic model validation, selection and evaluation. Various deep neural network models are designed and evaluated through multiple experiments.
Results suggest, that deep learning models, in fact, can forecast macroeconomic variables, preprocessing plays an important role, and automatic parameter estimation is a feasible approach for problems of large scale.
Initial Project Requirements
The essence of the project is to apply a machine learning based approach to acquiring empirical forecasting models of macroeconomic variables, such as inflation, unemployment and GDP for one or more economies (e.g. UK, EU, US). The project will use Keras, a high-level neural networks API, written in Python and capable of running on top of either TensorFlow, CNTK or Theano. It was developed with a focus on enabling fast experimentation. The library supports both convolutional networks and recurrent networks, as well as combinations of the two, and it runs seamlessly on CPU and GPU.
An MEng student will be expected to produce models from data that represents more than one economy, using at least 2 macro-parameters in each case. The benefits of including key commodities, such as oil and gold, should also be at least discussed in the design section (if adequate data cannot be gathered). A certain level of automation of the experimental setup is expected to be developed, and discussed in the report.