A problem with the one step ahead prediction is the trade-off between the noise cancelation and the details detection. The methods proposed up to now perform a global analysis of the data and lead to missing some details or at the opposite including noise. To solve this problem, we propose a preprocessing procedure inspired by the wavelet transform. This one decomposes the signal into some filtered series. Each of them is then modeled by a neural network which provides a one step prediction rule. The sum of these predictions can be seen as a prediction of the original series. Tests on sunspot data and comparison substantiate our approach.
展开▼
