The main purpose of this study is to introduce Bayesian neural
network approach for forecasting autoregressive moving average time
series models. The proposed approach is compared with the classical
Bayesian approach. To achieve these objectives first 48000 samples are
generated from different autoregressive moving average models, different
sample sizes (25, 50,100) were used for the network training and testing
.Then accuracy of the proposed approach was evaluated and compared
with the classical Bayesian method using three tools; the mean square
error , the mean absolute deviation of error and mean absolute error
ratio.Second both Bayesian neural network and the classical Bayesian
approaches were used to forecast five real data series according to three
cases: perfect prior, some prior and no prior .Two macro computer
programs were designed (MATLAB code). The first for Bayesian neural
network training, testing and comparing with Bayesian method, and the
second for calculating automatically the proposed neural network
forecasts.
Bayesian analysis of autoregressive moving average models is
difficult since the likelihood function is analytically intractable, which
causes problems in prior specification and posterior analysis.
The results showed that the performance of proposed Bayesian
neural network approach for forecasting is better than the performance of
Bayesian method. |
