Wavelet-based Forecasting of Short and Long Memory Time Series

A wavelet-based forecasting method for time series is introduced. It is based on a multiple resolution decomposition of the signal, using the redundant “à trous” wavelet transform which has the advantage of being shift-invariant. The result is a decomposition of the signal into a range of frequency scales. The prediction is based on a small number of coefficients on each of these scales. In its simplest form it is a linear prediction based on a wavelet transform of the signal. This method uses sparse modelling, but can be based on coefficients that are summaries or characteristics of large parts of the signal. The lower level of the decomposition can capture the longrange dependencies with only a few coefficients, while the higher levels capture the usual short-term dependencies. We show the convergence of the method towards the optimal prediction in the autoregressive case. The method works well, as shown in simulation studies, and studies involving financial data. Index Terms Wavelet transform, forecasting, resolution, scale, autoregression, time series, model