Wyniki wyszukiwania - BazTech

Ograniczanie wyników

1 Journal of Applied Computer Science

1 2007

Znaleziono wyników: 1

Liczba wyników na stronie

Wyniki wyszukiwania

Sortuj według:

Ogranicz wyniki do:

Dynamic Correlation Approach to Early Stopping in Neural Forecasting of Macroeconomic Indices

Michalak K., Raciborski R.

Journal of Applied Computer Science

2007

Vol. 15, nr 2

27-40

Neural networks are commonly used for modelling he behaviour of macroeconomic indices. One of the issues that arise in neural networks applications is that when a neural network is trained for too long the quality of the predictions tends to drop with the increasing number of training iterations. To overcome this problem various methods of early stopping are employed. In this paper an early stopping method based on dynamic correlation between time series introduced in our previous work is validated on macroeconomic indices. In correlation-based approach the decision whether to stop training or not is based on the prediction error value calculated for the time series that has possibly the lowest mean dynamic correlation with the time series being predicted. Experimental results show advantages of dynamic correlation method in predicting the behaviour of industrial production indices for EU-25 member countries. Experiments show that correlation-based method produces lower forecast errors compared to the early stopping performed using a subset of historical samples from the same time series. The results obtained imply that this observation is of a high statistical significance.