A. Elías Fernández, R. Jiménez Recaredo
Our approach relies on the selection of past observations that make the last function a central datum in the observed domain. The subsample projected ahead in time carries information about the temporal dependency that allows to provide point and band forecast.
The method is intuitive and easy to implement, then, may be attractive to a broad audience. Its applicability is enhanced by the proposed automatic parameter selection that makes the method completely data driven. Its usefulness is demonstrated in a simulation setting with various data generation processes and in real case studies. Electricity demand and monoxide emissions are considered, obtaining accurate point forecasts and narrow prediction bands, for a given nominal coverage level.
We can update predictions when the most recent observation is not completely observed in the full domain and the analysis of the subsamples provide insights about the drivers of the results and about the temporal structure.
Keywords: functional time series, band depth, central regions, prediction bands, forecasting, periodically correlated process, functional autoregressive processes, electricity demand, monoxide emissions.
Scheduled
GT6-3 Functional Data Analysis
September 5, 2019 2:45 PM
I3L9. Georgina Blanes building