Abstract:This article proposes a new approach to modeling high-dimensional time series by treating a $p$-dimensional time series as a nonsingular linear transformation of certain common factors and idiosyncratic components. Unlike the approximate factor models, we assume that the factors capture all the non-trivial dynamics of the data, but the cross-sectional dependence may be explained by both the factors and the idiosyncratic components. Under the proposed model, (a) the factor process is dynamically dependent and the idiosyncratic component is a white noise process, and (b) the largest eigenvalues of the covariance matrix of the idiosyncratic components may diverge to infinity as the dimension $p$ increases. We propose a white noise testing procedure for high-dimensional time series to determine the number of white noise components and, hence, the number of common factors, and introduce a projected Principal Component Analysis (PCA) to eliminate the diverging effect of the idiosyncratic noises. Asymptotic properties of the proposed method are established for both fixed $p$ and diverging $p$ as the sample size $n$ increases to infinity. We use both simulated data and real examples to assess the performance of the proposed method. We also compare our method with two commonly used methods in the literature concerning the forecastability of the extracted factors and find that the proposed approach not only provides interpretable results, but also performs well in out-of-sample forecasting. Supplementary materials of the article are available online.

Principal Dynamical Components

Dynamic Principal Component Analysis in High Dimensions

Data-driven Dimension Reduction for High-Dimensional Random Vibration Systems with Cubic Nonlinearity

Dynamic Principal Subspaces in High Dimensions

Reducing the Dynamic State Index to its main information using Principal Component Analysis

Principal Fitted Components for Dimension Reduction in Regression

Generalized functional dynamic principal component analysis

Modeling Multivariate Spatial-Temporal Data with Latent Low-Dimensional Dynamics

Modeling High-Dimensional Time Series: A Factor Model with Dynamically Dependent Factors and Diverging Eigenvalues

Dimensionality Reduction of Collective Motion by Principal Manifolds

Hierarchical disjoint principal component analysis

Principal Component Analysis using Frequency Components of Multivariate Time Series

Principal component analysis balancing prediction and approximation accuracy for spatial data

Fault detection approach based on linear dynamical system in principal component space

A Improved Common Principal Components Based Dimension Reduction Method for Multivariate Time Series Analysis

DYNAMIC PRINCIPAL COMPONENT REGRESSION: APPLICATION TO AGE-SPECIFIC MORTALITY FORECASTING

Dimension Reduction for High Dimensional Vector Autoregressive Models

Dimensionality Reduction in Discrete-Time Dynamical Systems

Dynamic Principal Subspaces with Sparsity in High Dimensions

Dynamic Processes Modeling and Monitoring Based on a Novel Dynamic Latent Variable Model.