Modeling Time Series of Real Systems using Genetic Programming

TL;DR

This paper demonstrates how Genetic Programming can effectively model and predict both simulated and real-world time series data, including plasma measurements and financial indices, with good accuracy.

Contribution

It introduces a GP-based framework for modeling diverse time series, showing its ability to produce accurate and generalizable models for complex systems.

Findings

GP models fit both simulated and real data well

Predictions range from very good to fair accuracy

Models generalize beyond the training data

Abstract

Analytic models of two computer generated time series (Logistic map and Rossler system) and two real time series (ion saturation current in Aditya Tokamak plasma and NASDAQ composite index) are constructed using Genetic Programming (GP) framework. In each case, the optimal map that results from fitting part of the data set also provides a very good description of rest of the data. Predictions made using the map iteratively range from being very good to fair.