| Summary: | Time series analyses in financial area have been attract some special attention in the recent years. The stock markets are examples of systems with a complex behaviour and, sometimes, forecasting a financial time series can be a hard task. In this thesis we compare linear against non-linear models, ARIMA and Artificial Neural Networks. Using the log returns of nine countries we tried to demonstrate that neural networks can be used to uncover the non-linearity that exists in the financial field. First we followed a traditional approach by analysing the characteristics of the nine stock series and some typical features. We also produce a BDS test to investigate the nonlinearity, the results were as expected, and none of the markets exhibit a linear dependence. In consequence, traditional linear models may not produce reliable forecasts. However, this didn’t mean that neural networks can. We trained four types of neural networks for the nine stock markets and the results between them were quite similar varying most in their structure and suggesting that more studies about the hidden units between the input and output layer need to be done. This study stresses the importance of taking into account nonlinear effects that are quite evident in the stock market MODELS.
|
|---|