next up previous

FROM NONLINEARITY TO PREDICTABILITY

Milan Palus
Institute of Computer Science, Academy of Sciences of the Czech Republic
Pod vodárenskou vezí 2, 182 07 Prague 8, Czech Republic
E-mail: mp@uivt.cas.cz, mp@santafe.edu

Dagmar Novotna
Institute of Atmospheric Physics, Academy of Sciences of the Czech Republic
Bocni II/1401, 141 31 Prague 4, Czech Republic
E-mail: nov@ufa.cas.cz

Abstract:

Detection of nonlinearity in experimental time series is usually based on rejection of a linear null hypothesis by a statistical test. Typically, the null hypothesis is a Gaussian process or a Gaussian process passed through a static nonlinear transformation or a similar simple alternative. Rejection of such a null is frequently interpreted as a detection of a deterministic nonlinear relation in data under study, which is, however, only one of possible alternatives. We show how variable variance, or seasonality in variance could lead to spurious identification of deterministic nonlinearity and discuss how to distinguish actual determinism in studied time series.



Neural Network World Vol. 7 No. 4,5 (1997) 451-460



Milan Palus 1998