Time Series Prediction with Direct and Recurrent Neural Networks

Authors : Lidio Mauro Lima DE CAMPOS

Pages : 7-15

View : 11 | Download : 9

Publication Date : 2017-08-23

Article Type : Research Paper

Abstract :Presents   a   comparative study for  prediction  of   time series   of the   Consumer   Price Index - CPI   using   recurrent   neural   network   insert ignore into journalissuearticles values( RNN);.   For this ,   three   models   are   designed   for networks   with   recurrent   and are   given   the changes   in `backpropagation` to allow them to incorporate the models ARX insert ignore into journalissuearticles values(Auto-Regressive with external input); and NARX insert ignore into journalissuearticles values(Nonlinear  Auto Regressive  with external input);. Furthermore, we present a third architecture, re-fed with the hidden layer, nicknamed ARXI, which is a special case of the Elman Network. Is carried out training for all networks and tests the ability to generalize them insert ignore into journalissuearticles values(identification stage);, in order to select the best architectures of recurrent networks to prediction of the IPC. After this stage, it makes the models validation, by means of the test the extrapolation capacity of the networks, i.e., presented data were not used during the training phase and gets the responses that indicate the capacity to predict future CPI for various times insert ignore into journalissuearticles values(validation phase);.   We conclude   that   NARX   networks   are   those with   best   performance and   that the   hybrid system   proposed by   [ 5]   constitutes   an   excellent   tool   when you   want   to   get   minimal networks   that make   a   series   of   perdition   satisfactorily .
Keywords : Recurrent Neural Networks, Time Series Prediction, NARX network