Neural networks and econometric methodologies for South African exchange rate forecasting

One of the most popular applications of neural networks is the prediction of financial data, in particular the prediction of exchange rates. Though promising results have been achieved, the performance of neural networks is hardly ever related to the performance of econometric methodologies. This paper therefore compares multiple linear regression (MLR), variable parameter regression (VPR), and fully connected single middle layer artificial neural network (ANN) models. A hybrid approach is used to specify the topology of the neural net. It is further believed that neural networks are capable of dealing with the problem of economic structural change. A comparison concerning the out-of-sample forecasting performance of monthly Rand/US$ exchange rate models, under conditions of structural instability, is therefore conducted. It is shown that ANN models outperform both the MLR and VPR models in terms of out-of-sample forecasting accuracy.