WAVELET NETWORKS WITH APPLICATIONS TO THE PREDICTION OF FINANCIAL SERIES
WAVELET NETWORKS WITH APPLICATIONS TO THE PREDICTION OF FINANCIAL SERIES

Summary/Abstract: Wavelet networks are special cases of neural networks and have become popular after the work of Zhang and Benveniste. Like neural networks, wavelet networks compute a linear combination of nonlinear functions whose form depends on adjustable parameters. However, while in neural networks the nonlinear dynamics are approximated by the superposition of sigmoid functions, in the wavelet networks these nonlinear dynamics are approximated by wavelet functions. As opposite to the activation functions (usually, sigmoid-shaped) of the neural networks, the wavelets have a rapid decay and tend to zero. The idea behind wavelet functions was to construct a transformation for the study of signals more interesting that the Fourier transform. The adjustment of the parameters is done by a learning procedure that uses input-output datasets to adapt their predicted responses in such a way to conform to the desired responses. Parameter estimation is an essential component in adaptive methods of estimation or prediction in general. The question is to find the parameters so that the model reproduces in an optimal way the observations. Both networks have the property of acting as universal approximators. This paper aims at using wavelet networks to predict financial time series with nonlinear dynamics.

• Details
• Contents