Abstract:
Time series forecasting (TSF) has been widely used in many application areas such as science, engineering,
and nance. Usually the characteristics of phenomenon generating a series are unknown
and the information available for forecasting is limited to the past values of the series. It is, therefore,
important to use an appropriate number of past values, termed lag, for forecasting. Although
ensembles (combining several learning machines) have been widely used for classi cation problems,
there is only a handful work for TSF problems. Existing algorithms for TSF construct ensembles by
combining base predictors involving di erent training parameters or data sets . The idea of ensemble
is also employed to nd the optimal parameter of predictors used for TSF. The aim of using di erent
parameters or data sets is to maintain diversity among the learning machines in an ensemble. It
has been known that the performance of ensembles greatly depends not only on diversity but also
on accuracy of the learning machines. However, the issue of accuracy is totally ignored in ensemble
approaches used for forecasting.
This thesis proposes a layered ensemble architecture (LEA) for TSF. Our LEA is consisted of two
layers. Each of the layers uses a neural network ensemble. However, tasks of ensembles in the two
layers are di erent. While the ensemble of the rst layer tries to nd an appropriate time window of
a given time series, it of the second layer makes prediction using the time window obtained from the
lower ( rst) layer. For maintaining diversity, LEA uses a di erent training set for each network in the
ensemble of the rst and second layers. LEA has been tested extensively on the time series data sets
of NN3 competition. In terms of prediction accuracy, our experimental results have showed clearly
that LEA is better than other ensemble and nonensemble algorithms.
