Time series forecasting for non-static environments: The DyFor genetic program model.

摘要

Current methods of time series forecasting assume a static environment, that is they assume the underlying data generating process is constant. Many time series encountered in real-world circumstances are produced in non-static environments where the underlying data generating process varies over time. When such time series are to be forecasted, current methods rely, to some degree, on human judgment to determine which historical data are valid for the existing environment and should be utilized for analysis. In many cases the time series in question is not well-understood and, thus, the forecasting task proves to be unwieldy.; Genetic Programming (GP) is an automatic programming technique that uses the Darwinian principle of survival of the fittest and sexual recombination (crossover) to evolve computer programs that solve (or approximately solve) problems. Several studies have applied GP to the task of forecasting with favorable results. However, these studies, like those applying other techniques, have assumed a static environment, making them unsuitable for many real-world time series which are generated by a varying process.; This study investigates the development of a new "dynamic" GP model that is specifically tailored for forecasting in non-static environments. This Dynamic Forecasting Genetic Program (DyFor GP) model incorporates methods to adapt to changing environments automatically as well as retain adaptations learned from previously encountered environments. Such past-learned adaptations prove useful when current environmental conditions resemble those of a prior setting. Specifically, these past adaptations allow for faster convergence to current conditions by giving the model searching process a "head-start" (i.e., by using learned knowledge to narrow the model search space).; The DyFor GP model is realized and tested for forecasting efficacy on real-world economic time series, namely the U.S. Gross Domestic Product and Consumer Price Index Inflation. Results show that the DyFor GP model outperforms benchmark models from leading studies for both experiments. These findings affirm the DyFor GP's potential as an adaptive, non-linear model for real-world forecasting applications and suggest further investigation.