Applying articial neural network techniques to the ARMA model

ARMA models and Artificial Neural Networks are commonly used approaches for forecasting timeseries. ARMA models are accurate and efficient, but difficult to use and inflexible to implement. Artificial Neural Networks are less efficient, but more flexible and easier to train. This thesis proposes a novel neural network, the ARMA-ANN, which is mathematically equivalent to the ARMA model, but trainable with commonly used neural network techniques. The ARMA-ANN was developed through a series of exploratory experiments, which evaluated different structures and training methods. It was tested against the forecasting ability of the ARMA model, producing results that were equivalent or better, especially in the case of larger models and real-time training. (note: 'ARMA' should be pronounced as a word phonetically, while 'ANN' is an acronym and should be pronounced as the three letters).