On projection methods for functional time series forecasting

Two nonparametric methods are presented for forecasting functional time series (FTS). The FTS we observe is a curve at discrete-time point. We address both one-step-ahead and dynamic updating. Dynamic updating forward prediction of the unobserved segment most recent curve. Among two proposed methods, first one straightforward adaptation to k -nearest neighbors univariate forecasting. second based on selection curves, termed envelope, that aims be representative in shape magnitude observation, either whole or observed part partially In similar fashion other projection successfully used forecasting, “project” curves envelope doing so, keep track next period evolution curves. applied simulated data, daily electricity demand, NOx emissions provide competitive results with often superior several benchmark predictions. approach offers model-free alternative statistical modeling study cyclic seasonal behavior many FTS.