Abstract:
The modelling technique known as Artificial Neural Networks (ANNs) is investigated. ANNs have the ability to detect and project non-linear relationships between variables. Further, they can adapt in dynamically changing environments while providing accurate results. A method of constructing ANNs in order to form a forecasting system is presented here. Further, in many of the applications studies, ANNs are fitted using crude guesses as to the efficient input parameters. In this study detailed investigations into parameter estimates are performed. In addition, ANNs and traditional models (ARIMA, seasonal smoothing, geometric Brownian motion, etc.) are constructed to forecast monthly inflation and the average monthly return on the money, bond and equity markets in South Africa from 1975 to 2010. The ANNs constructed are done through an integrated and isolated approach. The performance of the traditional and ANN models are compared. No general conclusion, as to which model is superior for all the applications considered, can be made. This suggests that ANNs perform as well as traditional models when forecasting financial markets. Further, it is found that the money market and inflation are forecast efficiently through all the models, over a single month. As the forecast period extends to three months the money market favours the traditional model. However, a forecast period of twelve months leads to the preference of ANNs in the case of the money market. Neither technique can forecast the equity or bond market accurately, as these require additional explanatory variables to those considered. As the forecast period increased, the forecast accuracy decreased for all the models. The integrated ANNs, which allow interaction between the markets, do not lead to improved forecasts which indicates that the relationships between the markets have a limited effect on the future values of the markets. Hybrid models are constructed, trained and tested for the money market and inflation. They are found to add value to traditional models when forecasting inflation but not the money market. The sensitivity of the performance of ANNs and the traditional model to different subsets of the inflation data is tested. No statistical difference between the models is found. The implementation advantages of ANNs are also
described.
