Functional Time Series Prediction

Ozan Sonmez

Stationary functional time series predictions via functional principal component analysis.

A new functional time series prediction method utilizing a decaying dimension of projec- tion space via functional principle components is proposed. Two step procedure is applied to obtain functional prediction where in the first step, the decay rate of the subspace di- mensions are chosen via minimizing an accumulated prediction error, and then in the second step the multivariate predictions are obtained using the dimensions chosen previously and the multivariate prediction is re-transformed into curves by Karhunen-Loeve expansion.