RETROSPECTIVE FORECASTS OF THE 2016 U.S. PRIMARY ELECTIONS An empirical comparison of evolutionary and gradient-based neural network training with applications in political forecasting Lennert Jansen 10488952
Abstract This thesis concerns an empirical comparison between differential evolution and gradient-based optimisation methods, applied to artificial neural network training. The gradient-based methods outperform differential evolution. A logistic regression model is considered. The results, however, suggest that a more elaborate network architecture is required to grasp the non-linearities in the data to the fullest extent.
Under the supervision of Dr. N.P.A. van Giersbergen Faculty of Economics and Business University of Amsterdam December 2017