Genetic Programming and Boosting Technique to Improve Time Series Forecasting
Luzia Vidal de Souza, Aurora Pozo, Cezar A. de F. Anselmo, Joel M. C. da Ros · Evolutionary Computation · 2009
In this paper we present the BCC algorithm that uses the correlation coefficients between the real and the forecasting value obtained using GP as a base learner. The correlation coefficient is used for update the weights and for the generation of the final formula. Differently from works found in the literature, in this paper we investigate the use of the correlation metrics as a factor, besides the error metric. This new approach, called Boosting using Correlation Coefficients (BCC), has been empirically obtained when trying to improve the results from the other methods. The correlation coefficient was considered because two algorithms could present the same mean error and different correlation coefficients for a dataset. This difference on behavior of the two algorithms can be measured by the correlation coefficient. A good correlation coefficient results in small errors for each example. The correlation coefficient is used in the proposed algorithm with two purposes: