Forecasting Chaotic Time Series: Global vs. Local Methods

We discuss the capabilities of global (neural network) and local (instance-based) methods for the dynamics reconstruction and forecasting of chaotic time-series. In particular, we investigate the performance of these methods as a function of the database length, with emphasis in the most frequent situation of a small to moderate number of registers available. Using the logistic map and the Mackey-Glass equation as examples, we conclude that with scarce data the neural network technique produces better results than a very eecient local method shown to outperform other algorithms in its class. However, for moderate computational time and/or medium-sized data sets the proposed local method can be highly competitive or even better than the global approach.