Consistent estimation of high-dimensional factor models when the factor number is over-estimated (Q2192324)

This paper refers to statistical inference on a model that has an increasing application especially in econometrics: the factor model in multidimensional time series. Specifically, a new method for estimating the number of relevant factors to be taken into account when proposing an adjusted model to a multidimensional time series. In addition to a new method for estimating the number of factors to be considered in the model, the authors propose a modified principal component (PC) estimator to address the problem of unreliability in the number of factor estimated. Let's see now a brief introduction to the mathematical aspects that define the problem faced in this paper. Let $(\Omega,F,P)$ be a probability space, $n\geq 2$, $t\geq 1$, and $r\geq 1$ integers. Let $X_{t}=(X_{1t},\dots,X_{nt})':\Omega\to\mathbb{R}^n$, $\varepsilon_{t}=(\varepsilon_{1t},\dots,\varepsilon_{nt})':\Omega\to\mathbb{R}^n$, $n$-dimensional random vectors, and $f_{t}=(f_{1t},\dots,f_{rt})':\Omega\to\mathbb{R}^{r}$ an $r$-dimensional random vector. Let $\Lambda_{n}=[\lambda_1,\dots,\lambda_{n}]'$ be a real $n\times r$ matrix with row vectors $\lambda_{i}=(\lambda_{i1},\dots,\lambda_{ir})$. It is assumed that \[X_{t} =\Lambda_{n}f_{t}+\varepsilon_{t},\tag{1}\] \[E(f_{jt}) = E(\varepsilon_{it})=0,\quad 1\leq j\leq r, 1\leq i\leq n, t\in\mathbb{N}.\] To identify the model are assumed: \begin{itemize} \item[i)] $E(f_{t}\cdot f_{t}')=Id_{r}$ (identity $r\times r$ matrix), $t\in\mathbb N$. \item[ii)] There exists a positive definite $r\times r$ matrix $H$ with distinct eigenvalues, such that \[ (1/n)\Lambda_{n}'\Lambda_{n}\xrightarrow[n\to\infty]{} H. \] \item[iii)] There exists $\lambda\in(0,+\infty)$ such that $\Vert\Lambda_{n}\Vert_{\max}\leq\lambda$ where \[ \Vert\Lambda_{n}\Vert_{\max}=\max_{1\leq j\leq r}\max_{1\leq i\leq n}|\lambda_{ij}|. \] \item[iv)] There exists $c_{\varepsilon}\in(0,+\infty)$ such that \[ \sum_{i=1}^n\sum_{i'=1}^na_{i}a_{i'}E(\varepsilon_{it}\cdot\varepsilon_{i't})<c_{\varepsilon} \] either $(a_{i})_{1\leq i\leq n}$ such that $\sum_{i=1}^na_{i}^2=1$. \item[v)] $E(f_{jt}\varepsilon_{it'})=0$, $1\leq i\leq n$, $1\leq j\leq r$, $t\text{ and }t'\in \mathbb N$. \end{itemize} In the financial mathematics literature, \[ X_{nt}=\Lambda_{n}f_{t}, n \text{ and }t\text{ positive integers} \] is the common component driven by the latent factors $f_{t}$, and $(\varepsilon_{it})_{1\leq i\leq n}$ is the idiosyncratic component of the model (1). For each $1\leq i\leq n$ and $1\leq t\leq T$, the classic PC estimator of $X_{it}=\lambda_{i}'f_{t}$ is \[ X_{it}^{pc}=\sum_{j=1}^{r}w_{X,ij}w_{X,j}'X_{t}, \] where $w_{X,j}'$ is the estimator of the \(j\)-th normalized eigenvector of the covariance matrix of $X_{t}$, and $r$ is an estimator of $r$, the number of factors to be considered. The eigenvalues of the matrices which appear are considered in descending order of magnitude. Under assumptions concerning common and idiosyncratic components the authors prove a certain form of asymptotic consistency of the $X_{it}^{pc}$ estimator. However, they have shown, in previous papers, that the most commonly used estimators of the number of factors to be considered overestimate such a number, leading to not negligible errors. Athough working with fewer than relevant factors can also lead to considerable errors. This paper, by means of a rigorous formal analysis, reinforces those results about the problems that arise when overestimating the number of factors. Therefore, a new PC estimator is proposed, called ``rescaled PC estimator'': \[ X_{it}^{sc} =\sum_{j=1}^{r}w_{X,ij}^{sc}(w_{X,j}^{sc})'X_{t}, \] \[ w_{X,j}^{sc} =\nu_{j}^{-1}w_{X,j},\quad \nu_{j}=\max\{1,((\sqrt n)/(c_{w})) \max_{1\leq i\leq n}|w_{X,ij}|\}, \] $c_{w}$ being a conveniently chosen constant (a guide on how to choose it is given in the paper). $\nu_{j}$ penalizes the contribution of possible spurious factors. The best performance of the new PC estimator compared to the classic PC estimators is shown through a simulation study, and with real financial series data. In the situations analyzed by simulation, the authors point out that the usual estimator of the covariance matrix based on factor analysis is not reliable when working with an overestimated number of factors. Also, this paper shows that the new proposed PC estimator with the number of factors estimated by the method proposed here has the same asymptotic convergence rate as the classical PC estimator when the number of factors is known. A Subsection of the paper details the notation used which provides clarity and understanding of the concepts and results exposed. Several results stated in this paper are proved in another paper by the same authors, which can be downloaded free of charge. The list of bibliographical references is extensive and up to date.