No label defined (Q4936192)

The author starts with a review of the most frequently used nonlinear models of time series (functional coefficient models, bilinear models, threshold AR models, smooth transition AR models, ARCH models, doubly stochastic models). Then tests for nonlinearity of time series are introduced. A simulation study confirms the known fact that the tests react differently to different nonlinear models. Moreover, some tests can lead to false detection of nonlinearity, although the model is linear but nonnormal. It is shown that some problems can be solved using bootstrap. Namely, the tests based on bootstrap are robust against deviations from normality. Also some methods for model identification can be based on bootstrap. Simulations confirm that it is possible to select in this way the true model from three most important classes of nonlinear time series. The described methods are applied to some sets of real data, which include IBM price stock series and Canadian lynx data.