A0447
Title: Imputation for tensor time series
Authors: Zetai Cen - London School of Economics and Political Science (United Kingdom) [presenting]
Clifford Lam - London School of Economics and Political Science (United Kingdom)
Abstract: It is prevalent to have missing data in different areas, such as econometrics, and imputation is one of the common approaches to deal with it. With the fast-growing data size and complicated data structure, tensor time series is considered, and an imputation procedure based on factor analysis is proposed. Other researchers generalised The idea previously, but more general data are allowed. Specifically, a re-arrangement algorithm is used to construct different blocks of tensor data, and previous work has been adopted to estimate the tensor factor model for imputation. Thus, weak factors and serial and cross-correlations in the idiosyncratic errors and time series variables are allowed with bounded fourth-order moments. Also, iterative imputation is performed by re-estimating the factor structure to improve the imputation result. Simulations under different settings are performed, with appealing results even with weak factors and heavy-tailed data, and the one-step iteration is good enough and enjoys a short imputation time.
