Forecasting Time Series with Neural Nets

FFNN (Feed-Forward Neural Nets) are one of the most widely used neural nets. In this thesis the FFNN architecture is examined and compared with statistical time series models for a variety of time series prediction problems.  FFNN do not assume any probability models, while statistical models are based on the probability model.  Therefore, if the goal of the modelling is rigorous quantification of uncertainty, statistical models are more suitable.  However if the goal is merely prediction, we demonstrate that neural nets have a lot to offer.  Widely different parameter settings in the neural net approach often lead to models which make virtually the same predictions. Neural net offer a more flexible approach to model building which is especially helpful in nonlinear and nonGaussian situations. In this thesis, the performance by NN models and statistical models for prediction is examined by using visualization techniques and statistical tests.

Keywords:
Feed-Forward Neural Nets,
Linear and nonlinear time series models,
Forecasting,
Nonlinear time series,
Visualization