The aim of this paper is to define efficient strategies for the modeling an
d predicting of short and noisy time series with neural networks. Several c
omplementary methods are tested on short series constructed from the Lorenz
system which has been spoiled with various levels of measurement or dynami
cal noise. The best strategies are selected from the simulation results acc
ording to the level and the noise characteristics. In the presence of measu
rement noise we show that over-sizing of the embedding dimension of the lea
rning set increases the powerness of neural network fits. In the case of dy
namical noise spoiling, we found that generation of a new trajectory predic
ted with local operators, amplifying information of the original series, al
lows the usage of neural networks as in the case of measurement noise, and
so, avoids over-fitting problems possible with very short series. The strat
egies applied to the real biological and astronomical data (whooping cough
in Great Britain and Wolf sunspot numbers) revealed their deterministic ske
letons showing chaotic attractors. (C) 2001 Elsevier Science Ltd. All right
s reserved.