Time series modeling and Forecasting using genetic algorithms

Abstract

The time series forecast is a very complex problem, consisting in predicting the behavior of a data series with only the information of the previous sequence. In this thesis, a highly comparative framework for time-series modeling and forecasting is developed to give a good models for time series to perform the prediction as well as linear predictors. For this, we are reduce errors of time series models, by used genetic algorithms GAs, one of robustness methods of optimization. GAs is inspired by natural evolution theories, apply operations of reproduction, crossover and mutation to candidate solutions according to their relative fitness scores in the successive populations of candidates Holland (1975). We choose the mean square errors MSE and Akaike criteria information AIC the objectives functions for optimization to calculate there minimums for select the optimal model. This method is based on the evolution of set of rules genetically codified, two types of coding GAs are used, the binary coded GAs (BCGA) and reel coded GAs (RCGA). In the first time we optimize MSE and AIC with BCGA, and in the second time with RCGA. By order to determine the best method, we compar between different methodologies and we contribute in this comparaive study by use a modification to errors of model given by GAs by normalized them, for have a better model and perform the forecast. The computer simulation results obtained demonstrate that GAs have the potential to become a powerful tool for time series modeling and forecasting and as well as when we use the advanced GAs. To illustrate our studies, we supports all chapters by some examples application on real time series data.