A0497
Title: Spatio-temporal forecast by spatial VAR with co-integration and functional PCA on empirical dynamic quantile series
Authors: Guoqi Qian - The University of Melbourne (Australia) [presenting]
Xianbin Cao - The University of Melbourne (Australia)
Abstract: A spatial auto-regression co-integration (SARC) model, combined with techniques of empirical dynamic quantile (EDQ) series and functional principal component analysis (FPCA), is proposed to analyze time-series data observed from multiple spatial locations and to provide spatio-temporal forecast. The data in consideration are often temporally non-stationary, spatially correlated, and high-dimensional in both time and space, posing statistical and computational challenges in analyzing them. Using the proposed method, we tackle these challenges by first applying the SARC model to a small number of EDQ series of the data so that the residual vector time series from the model become stationary and uncorrelated. We then apply FPCA to these residual vector series to forecast the EDQ values at a future time. Finally, we extrapolate the EDQ series forecasts to the whole space domain according to the estimated quantile levels of the corresponding spatial locations, resulting in further spatio-temporal forecasts in the intended space-time domain. The proposed SARC-EDQ-FPCA method is evaluated using simulated data before being applied to analyze a real spatio-temporal data set on a landslide.
