Statistical evaluation of symbolic regression forecasting of time-series

摘要

This is an evaluation of the ability of symbolic regression to predict time-series. Symbolic regression is an application of genetic programming. Three codes-GPCPP, GPQuick, and Vienna University GP Kernel - written in C++ were tested. Six models generated data by linear, nonlinear, and pseudo-random processes, and the three codes were employed to search for the six data generating processes. The results suggest that: (1) complexity and predictability are inversely related, (2) the symbolic regression technique is successful in predicting less complex processes, and (3) all three failed to find a data generating process for pseudo-random data.