Predicting timeseries of chaotic systems can be a very difficult task. Most methods employed for such a feat are typically relying on large neural networks and machine learning. One does not need them though! In the package
TimeseriesPrediction we are presenting methods that instead take advantage of dynamical systems theory. Many such methods exist, like for example Cluster weighted modelling or networks of dynamical systems. The first method included in
TimeseriesPrediction, which is also simplest one, is called "local modelling".
Local modelling predicts timeseries using a delay embedded state space reconstruction. It finds the nearest neighbors of a query point within this reconstructed space and applies a local model to make a prediction. "Local" model refers to the fact that the images (future points) of the neighborhood of a point are the only component used to make a prediction.
In contrast to typical neural networks applications, there is no training happening in this approach. A given timeseries dataset constitutes a pool of points one uses to make predictions from.
TimeseriesPrediction has two local models (usable in any prediction scheme). See
AbstractLocalModel for more details.
One of the biggest strengths of
TimeseriesPrediction is a robust, simple, and feature-rich interface that can predict the evolution of spatio-temporal systems (commonly represented by PDEs or by "map lattices" (coupled maps)). To see the full interface please visit the spatio-temporal prediction page. In addition, the spatio-temporal examples page is full of runnable code that displays the capabilities of spatio-temporal prediction of