Different neural networks and regression techniques have been compared in a time series forecasting task. We have shown that the important differences in prediction accurracy, computational cost and additional information provided by each method are important factors to be considered for a specific application. These factors are also function not only of the amount of data used but also of noise.
Bayesian neural networks using Markov chain Monte Carlo simulations
perform better than other NN techniques on the Mackey Glass problem,
only outperformed by MARS on the clean data set. Their relative computational costs are high but
they do provide prediction intervals which may be useful in certain
applications.
Also, in real world problems with noisy and scarce data, the fact that
Bayesian neural networks do not overfit could be of practical value.
Further study is being done adding new techniques to the benchmark
trying them on different data sets and estimating formalized prediction intervals.
.