| dc.description.abstract | Poor forecasting performance and, sometimes, dubious results leads the financial time series researcher to ask if is it possible to really forecast anything from these series. Thus, it is of great relevance the study of the predictability of this kind of data. If it is, somehow, known that these series are to be unpredictable, then there is no point in trying to forecast such series. Some statistical tools for the detection of time dependencies are presented and used on stock market series. Measures based on the estimation of dimension, entropy and information exchange are used as a way of discriminating between these series and some known models (e.g., white gaussian noise, AR and GARCH models and the logistic map) in terms of complexity, regularity and information flow from past to present values of the series. These measures are used to test the series against the null hipotesis of independence and linear, nonlinear and exclusive heteroscedastic dependence. The discriminating power of these measures was obtained by the use of surrogate data analysis. The presence of heteroscedasticity was seen to have significant undesirable effects on the results. Also, the BDS test was used as a way to quantify the predictability, by considering only the correlations not due to heteroscedasticity. The ultimate test for predictability of a time series is its forecasting power. In this sense, the results of the predictability study were compared against the prediction errors obtained by RBF modelling of the data. It is shown that the results of both prediction and predictability are in agreement. | |
