EconBase
← Back to paper

Augmented Factor Models with Applications to Validating Market Risk Factors and Forecasting Bond Risk Premia

Extracted main text — title through conclusion, appendix excluded. This is what our citation measures are computed over, published so the extraction can be checked by eye.

101,582 characters · 26 sections · 40 citation commands

Rendered from LaTeX for readability, not typeset faithfully. Citation keys are highlighted; maths is left as source; figures, tables and equation environments are summarised rather than reproduced; unrecognised commands are greyed out so nothing is silently dropped. Email addresses are removed.

Augmented Factor Models with Applications to Validating Market Risk Factors and Forecasting Bond Risk Premia

\def\spacingset#1{ {#1}} \spacingset{1}

\if11 \fi

\if01 {

center[center omitted — 86 chars of source]

} \fi

abstractWe study factor models augmented by observed covariates that have explanatory powers on the unknown factors. In financial factor models, the unknown factors can be reasonably well explained by a few observable proxies, such as the Fama-French factors. In diffusion index forecasts, identified factors are strongly related to several directly measurable economic variables such as consumption-wealth variable, financial ratios, and term spread. With those covariates, both the factors and loadings are identifiable up to a rotation matrix even only with a finite dimension. To incorporate the explanatory power of these covariates, we propose a smoothed principal component analysis (PCA): (i) regress the data onto the observed covariates, and (ii) take the principal components of the fitted data to estimate the loadings and factors. This allows us to accurately estimate the percentage of both explained and unexplained components in factors and thus to assess the explanatory power of covariates. We show that both the estimated factors and loadings can be estimated with improved rates of convergence compared to the benchmark method. The degree of improvement depends on the strength of the signals, representing the explanatory power of the covariates on the factors. The proposed estimator is robust to possibly heavy-tailed distributions. We apply the model to forecast US bond risk premia, and find that the observed macroeconomic characteristics contain strong explanatory powers of the factors. The gain of forecast is more substantial when the characteristics are incorporated to estimate the common factors than directly used for forecasts.

{\it Keywords:} Heavy tails, Forecasts; Principal components; identification.

\spacingset{1.45}

Introduction

In this paper, we study the identification and estimations of factor models augmented by a set of additional covariates that are common to all individuals. Consider the following factor model:

equation[equation omitted — 113 chars of source]

Here $\bfm y_t=(y_{1t},..., y_{Nt})'$ is the multivariate outcome for the $t^{th}$ observation in the sample; $\bfm f_t$ is the $K$-dimensional vector of latent factors; $\ensuremath{\boldsymbol{\Lambda}}=(\ensuremath{\boldsymbol{\lambda}}_1,....,\ensuremath{\boldsymbol{\lambda}}_N)'$ is an $N\times K$ matrix of nonrandom factor loadings; $\bfm u_t=(u_{1t},...,u_{Nt})'$ denotes the vector of idiosyncratic errors. In addition to $\{\bfm y_t\}_{t=1}^T$, we also observe variables, denoted by $\bfm x_t$, that have some explanatory power on the unknown factors and hence impact on observed vector $\bfm y_t$. We model $\bfm f_t$ by using the model

equation[equation omitted — 70 chars of source]

for some (nonparametric) function $\bfm g = E(\bfm f_t|\bfm x_t)$. Here $\bfm g(\bfm x_t)$ is interpreted as the component of the factors that can be explained by the covariates, and $\bfsym \gamma_t$ is the components that cannot be explained by the covariates. We aim to provide an improved estimation procedure when the factors can be partially explained by several observed variables $\bfm x_t$. In addition, by accurately estimating $\bfsym \gamma_t$, we can estimate the percentage of both explained and unexplained components in the factors, which describes the proxy/explanatory power of covariates.

Note that model ((ref)) implies:

equation[equation omitted — 187 chars of source]

where $\operatorname{cov}(\bfm y_t)$ and $\operatorname{cov}(\bfm u_t)$ respectively denote the $N\times N$ variance-covariance matrices of $\bfm y_t$ and $\bfm u_t$; $\operatorname{cov}(\bfm f_t)$ denotes the $K\times K$ variance-covariance matrix of $\bfm f_t$. Under usual factor models without covariates, $\frac{1}{\sqrt{N}}\ensuremath{\boldsymbol{\Lambda}}$ is identified asymptotically as the first $K$ eigenvectors of $\operatorname{cov}(\bfm y_t)$ as $N\to\infty$ and can be estimated using the first $K$ eigenvectors of the sample covariance matrix of $\bfm y_t$ (e.g,, SW02, bai03).

commentApparently, $\ensuremath{\boldsymbol{\Lambda}} \operatorname{cov}(\bfm f_t)\ensuremath{\boldsymbol{\Lambda}}' $ is a low-rank matrix so long as $K<N$. Since $\{\bfm f_t, \bfm u_t\}_{t\leq T}$ is unobservable, what makes the decomposition ((ref)) identifiable is the following four typical conditions (they are not necessary conditions): (i) $N\to\infty$. (ii) $\{\nu_{y,1}>...>\nu_{y,K}\} $ should be distinct, representing the top $K$ eigenvalues of $\operatorname{cov}(\bfm y_t)$. (iii) All the eigenvalues of the $N\times N$ matrix $\operatorname{cov}(\bfm u_t)$ are bounded from above by a constant $c_u>0.$ (iv) All the eigenvalues of the $K\times K$ matrices $\ensuremath{\boldsymbol{\Lambda}}'\ensuremath{\boldsymbol{\Lambda}}/ N$ and $\operatorname{cov}(\bfm f_t)$ are bounded from below by a constant $c_\Lambda>0$. To see how this is helpful to identify $\ensuremath{\boldsymbol{\Lambda}}$, for simplicity of our explanations, we temporarily assume $\ensuremath{\boldsymbol{\Lambda}}'\ensuremath{\boldsymbol{\Lambda}}/N=\bfm I_K$, and that $\operatorname{cov}(\bfm f_t)$ is a diagonal matrix. Then, it is easy to see that the diagonal entries of $\operatorname{cov}(\bfm f_t)$ are the first $K$ eigenvalues of $\frac{1}{N} (\operatorname{cov}(\bfm y_t) -\operatorname{cov}(\bfm u_t) )$, with the columns of $\frac{1}{\sqrt{N}}\ensuremath{\boldsymbol{\Lambda}}$ the corresponding eigenvectors. Regarding $\operatorname{cov}(\bfm u_t)$ as the perturbation, let $\ensuremath{\boldsymbol{\Xi}}_y$ be an $N\times K$ matrix whose columns are the eigenvectors corresponding to the first $K$ eigenvalues of $\operatorname{cov}(\bfm y_t)$. We apply the Davis-Kahan $\sin\theta$-theorem to reach: $$ \|\frac{1}{\sqrt{N}}\ensuremath{\boldsymbol{\Lambda}}-\ensuremath{\boldsymbol{\Xi}}_y\| = O(N^{-1}). $$ Thus, $\frac{1}{\sqrt{N}}\ensuremath{\boldsymbol{\Lambda}}$ is identified as the first $K$ eigenvectors of $\operatorname{cov}(\bfm y_t)$ asymptotically and can be estimated using the first $K$ eigenvectors of the sample covariance matrix of $\bfm y_t$ (e.g,, SW02, bai03). In the general setting that allows $\ensuremath{\boldsymbol{\Lambda}}'\ensuremath{\boldsymbol{\Lambda}}/N\neq \bfm I_K$, and non-diagonal $\operatorname{cov}(\bfm f_t)$, the above arguments still hold, but $\ensuremath{\boldsymbol{\Lambda}}$ is identified up to a rotation transformation. We summarize the result as follows. \begin{prop} Suppose (i)-(iv) hold. Consider a general setting in which $\ensuremath{\boldsymbol{\Lambda}}'\ensuremath{\boldsymbol{\Lambda}}/N$ is not necessarily $\bfm I_K$ and $\operatorname{cov}(\bfm f_t)$ is not necessarily diagonal. Then there is a $K\times K$ matrix $\bfm H_y $ so that $$ \|\frac{1}{\sqrt{N}}\ensuremath{\boldsymbol{\Lambda}}\bfm H_y-\ensuremath{\boldsymbol{\Xi}}_y\| = O(N^{-1}). $$ \end{prop} \begin{rem} Both (iii) (iv) can be weakened to allow the eigenvalues of $\ensuremath{\boldsymbol{\Lambda}}'\ensuremath{\boldsymbol{\Lambda}}/N$ to decay slowly, and that the eigenvalues of $\operatorname{cov}(\bfm u_t)$ to grow slowly. But this would slow down the identification rate in Proposition (ref). \end{rem} It is important to note that this identification is only asymptotic (requires $N\to\infty$), in order to remove the effect of $\operatorname{cov}(\bfm u_t)$. In contrast, we show that with the additional common covariates and a rank condition, conditions (i)- (iii) can be removed, and condition (iv) is modified. The “exact identification" can be achieved, for any finite $N>K$. Neither do we need $N\to\infty$ to estimate $(\ensuremath{\boldsymbol{\Lambda}}, \bfm g(\bfm x_t))$. As we shall show in this paper, $N\to\infty$ is only required to improve the estimation accuracy for $\bfsym \gamma_t$.

With additional covariates, on the other hand, exact identification can be achieved through covariance of the “smoothed data”. By (ref), assuming exogeneity of $\bfm x_t$, we have $ E(\bfm y_t|\bfm x_t) = \ensuremath{\boldsymbol{\Lambda}} E(\bfm f_t | \bfm x_t) $ so that it becomes a “noiseless" factor model with smoothed data $ E(\bfm y_t|\bfm x_t)$ as input and $E(\bfm f_t | \bfm x_t)$ as latent factors. The factor loadings and latent factors can be extracted from

equation[equation omitted — 94 chars of source]

It is easy to see from the model that

equation[equation omitted — 136 chars of source]

where $\bfsym \Sigma_{f|x} = E\{E(\bfm f_t|\bfm x_t)E(\bfm f_t|\bfm x_t)'\}$ is a $K\times K$ low-dimensional positive definite matrix. This decomposition is to be compared with ((ref)), where the noise covariance $\operatorname{cov}(\bfm u_t)$ removed. Therefore, as long as $\bfsym \Sigma_{f|x}$ is of full rank, $\ensuremath{\boldsymbol{\Lambda}}$ falls in the eigenspace generated by $\bfsym \Sigma_{y|x}$. In other words, $\ensuremath{\boldsymbol{\Lambda}}$ is identifiable up to an orthogonal transformation. Because of such exact identification, we allow $N$ to be finite as a special case. The number of factors is assumed to be known throughout the paper. In practice, $K$ can be consistently estimated by many methods such as AIC, BIC-based criteria, or eigenvalue-ratio methods studied in LamYao,AH.

commentIn general, we define a spiked low-rank covariance as the follows: \begin{defn} Suppose $N>K$ and $N$ either grows or stays constant. An $N\times N$ “spiked low-rank matrix" is a matrix $\bfsym \Sigma$ of the following form: $$\bfsym \Sigma= \bfm A \ensuremath{\boldsymbol{\Xi}}\bfm A' , $$ where (a.i) $\bfm A$ is an $N\times K$ matrix, and $K$ stays constant. In addition, there are constants $\underline{c}_A, \bar{c}_A>0$ so that all the eigenvalues of $\frac{1}{N}\bfm A'\bfm A$ are in $[\underline{c}_A, \bar{c}_A]$. (a.ii) $\ensuremath{\boldsymbol{\Xi}}$ is a full rank $K\times K$ positive definite matrix. \end{defn} In addition, we require $\lambda_{\min}(\ensuremath{\boldsymbol{\Xi}})$ be either bounded away from zero or decay very slowly, without requiring a specific lower bound on $\lambda_{\min}(\ensuremath{\boldsymbol{\Xi}})$ in this definition. For such matrices $\bfsym \Sigma$, there are exactly $K$ number of nonzero eigenvalues, whose magnitudes are lower bounded by $$ \lambda_{\min}(\ensuremath{\boldsymbol{\Xi}}) \lambda_{\min}(\bfm A'\bfm A)\geq \underbar{c}_A N\lambda_{\min}(\ensuremath{\boldsymbol{\Xi}}). $$ When $N$ is very large and $\lambda_{\min}(\ensuremath{\boldsymbol{\Xi}})>c$ for some constant $c>0$, the nonzero eigenvalues of $\bfsym \Sigma$ can grow at rate $O(N)$, giving rise to the name “spiked" low-rank matrix. When either $N$ is finite or $\lambda_{\min}(\ensuremath{\boldsymbol{\Xi}})$ decays to zero, $N\lambda_{\min}(\ensuremath{\boldsymbol{\Xi}})$ does not need to grow, but we generally require $N\lambda_{\min}(\ensuremath{\boldsymbol{\Xi}})$ be larger than the statistical errors of estimating $\bfsym \Sigma$ in order to consistently estimate its leading eigenvectors. Set $\bfm A=\ensuremath{\boldsymbol{\Lambda}}$ and $\ensuremath{\boldsymbol{\Xi}}=\bfsym \Sigma_{f|x}$. Then by the above definition, $\bfsym \Sigma_{y|x}$ is a spiked low-rank matrix so long as $\ensuremath{\boldsymbol{\Lambda}}$ and $\bfsym \Sigma_{f|x}$ satisfy conditions $(a.i)$ and $(a.ii)$, and $\lambda_{\min}(\bfsym \Sigma_{f|x} )$ has a proper lower bound, representing the signal strength of the model. We provide conditions under which the columns of the loading matrix $\ensuremath{\boldsymbol{\Lambda}}$ are identified (up to a rotation) as the leading eigenvectors of $\bfsym \Sigma_{y|x}$ for each finite $N>K$, and construct PCA estimators based on the estimated low-rank matrix. It turns out, as we show in the paper, this leads to identifications with a finite $N$, and faster rates of convergence than the benchmark methods.

The above discussion prompts us the following new method to estimate the factor loadings $\ensuremath{\boldsymbol{\Lambda}}$ that incorporates the explanatory power of $\bfm x_t$: (See Section 3 for details of estimators)

(i) (robustly) regress $\{\bfm y_t\}$ on $\{\bfm x_t\}$ and obtain fitted value $\{\widehat\bfm y_t\}$;

(ii) conduct the principal components analysis (PCA) on the fitted data $(\widehat\bfm y_1,...,\widehat\bfm y_T)$ to estimate the factor loadings. \\ We employ a regression based on huber1964robust's robust M-estimation in step (i). The procedure involves a diverging truncation parameter, called adaptive Huber loss, to reduce the bias when the error distribution is asymmetric fan2017estimation. This allows our procedure to be applicable to data with heavy tails.\footnote{In this paper, by “heavy-tail" we mean tail distributions of $(\bfm u_t, \bfm y_t)$ that are heavier than the usual requirements on the high-dimensional factor model (which are either exponentially-tailed or have eighth or higher moments). But we do not allow large outliers on the covariates. }

There are two important quantities that determine the rates of convergence for the estimators: the “signal" $\bfsym \Sigma_{f|x}= E\{E(\bfm f_t|\bfm x_t)E(\bfm f_t|\bfm x_t)'\}$ and the “noise" $\operatorname{cov}(\bfsym \gamma_t)$. The rates of convergence are presented using these two quantities. Their relative strengths determine the rates of convergence of the estimated factors and loadings.

Under model ((ref)), we can test $\bfsym \gamma_t=0$ almost surely in the entire sampling period, under which the observed $\bfm x_t$ fully explain the true factors. This is the same as testing $$ H_0: \operatorname{cov}(\bfsym \gamma_t)=0. $$ While it is well known that the commonly used Fama-French factors have explanatory power for most of the variations of stock returns, it is questionable whether they fully explain the true (yet unknown) factors. These observed proxies are nevertheless used as the factors empirically, and the remaining components ($\bfsym \gamma_t$ and $\bfm u_t$) have all been mistakenly regarded as the idiosyncratic components. The proposed test provides a diagnostic tool for the specification of common factors in empirical studies, and is different from the “efficiency test" in the financial econometric literature (e.g., GRS, PY, gungor2013testing). While the efficiency test aims to test the asset pricing model through whether the alphas are zero for the specified factors, a rejection could be due to either mispecified factors or the existence of outperforming (underperforming) assets. In contrast, here we directly test whether the factor proxies are correctly specified. We test the specification of Fama French factors for the returns of S&P 500 constituents using rolling windows. We find that the null hypothesis is more often to be rejected using the daily data compared to the monthly data, due to a larger volatility of the unexplained factor components. The estimated overall volatility of factors varies over time and drops significantly during the acceptance period.

Further Literature

In empirical applications, researchers frequently encounter additional observable covariates that help explain the latent factors. In genomic studies, in the study of breast cancer data such as the Cancer Genome Atlas (TCGA) project cancer2012comprehensive, there are additional information of cancer subtype for each sample. These cancer subtypes can be regarded as a partial driver of the factors for gene expression data. In financial time series forecasts, researchers often collect additional variables that characterize financial markets. The Fama-French factors are well-known to be related to the factors that drive financial returns FF.

Most existing works simply treat $\bfm x_t$ as a set of additional regressors in ((ref)). This approach does not take advantage of the difference of observed variables (e.g. aggregated versus disaggregated macroeconomic variables; gene expressions versus clinical information) and the explanatory power of the covariates on the common factors, and hence does not lead to improved rates of convergence even if the signal is strong. The most related work is li2016supervised, who specified $\bfm f_t$ as a linear function of $\bfm x_t$. Also, huang2010combine proposed to use the estimated $\bfm g(\bfm x_t)$ to forecast. Moreover, our expansion is also connected to the literature on asymptotic Bahadur-type representations for robust M-estimators, see, for example, portnoy1985asymptotic, mammen1989asymptotics, among others.

The “asymptotic identification" was described perhaps first by CR. In addition, there has been a large literature on both the static and dynamic factor models, and we refer to Lawley, Forni05, SW02, BN02, bai03, doz, onatski, poet, among many others.

The rest of the paper is organized as follows. Section 2 establishes the new identification of factor models. Section (ref) formally defines our estimators and discusses possible alternatives. Section (ref) presents the rates of convergence. Section (ref) discusses the problem of testing the explanatory power. Section (ref) applies the model to forecasting the excess return of US government bonds. We present the extensive simulation studies in Section (ref) Finally Section (ref) concludes. The supplement also contains all the technical proofs.

Throughout the paper, we use $\lambda_{\min}(\bfm A)$ and $\lambda_{\max}(\bfm A)$ to denote the minimum and maximum eigenvalues of a matrix $\bfm A$. We define $\|\bfm A\|_F=\operatorname{tr}^{1/2}(\bfm A'\bfm A)$, $\|\bfm A\|=\lambda_{\max}^{1/2}(\bfm A'\bfm A)$, $\|\bfm A\|_1 =\max_{j} \sum_{i} |a_{ij}|$ and $\|\bfm A\|_{\max}=\max_{i,j}|a_{ij}|$. For two sequences, we write $a_T\gg b_T$ or $b_T\ll a_T$ if $b_T=o(a_T)$ and $a_T\asymp b_T$ if $a_T=O(b_T)$ and $b_T=O(a_T).$

Identification of the covariate-based factor models

Identification

Suppose that there is a fixed $d$-dimensional observable vector $\bfm x_t$ that is: (i) associated with the latent factors $\bfm f_t$, and (ii) mean-independent of the idiosyncratic term. Taking the conditional mean on both sides of ((ref)), we have

equation[equation omitted — 105 chars of source]

This implies

equation[equation omitted — 136 chars of source]

where

eqnarray*[eqnarray* omitted — 166 chars of source]

Note that $ E(\bfm y_t|\bfm x_t) $ is identified by the data generating process with observables $\{(\bfm y_t, \bfm x_t)\}_{t\leq T}$, but $\bfsym \Sigma_{f|x}$ is not because $\bfm f_t$ is not observable. Since $N>K$, ((ref)) implies that $\bfsym \Sigma_{y|x}$ is a low-rank matrix, whose rank is at most $K.$ Furthermore, we assume $ \bfsym \Sigma_{f|x}$ is also full rank, so $\bfsym \Sigma_{y|x}$ has exactly $K$ nonzero eigenvalues.

To see how the equality ((ref)) helps achieve the identification of $\ensuremath{\boldsymbol{\Lambda}}$ and $\bfm g(\bfm x_t)$, for the moment, suppose the following normalization holds:

equation[equation omitted — 180 chars of source]

Then right multiplying ((ref)) by $\ensuremath{\boldsymbol{\Lambda}}/N$, by the normalization condition, $$ \frac{1}{N}\bfsym \Sigma_{y|x}\ensuremath{\boldsymbol{\Lambda}} =\ensuremath{\boldsymbol{\Lambda}} \bfsym \Sigma_{f|x}. $$ We see that the ($K$) columns of $\frac{1}{\sqrt{N}}\ensuremath{\boldsymbol{\Lambda}}$ are the eigenvectors of $\bfsym \Sigma_{y|x}$, corresponding to its $K$ nonzero eigenvalues, which also equal to the diagonal entries of $ \bfsym \Sigma_{f|x} $. Furthermore, left multiplying $\ensuremath{\boldsymbol{\Lambda}}'/N$ on both sides of ((ref)), one can see that even if $\bfm f_t$ is not observable, $E(\bfm f_t|\bfm x_t)$ is also identified as: $$ \bfm g(\bfm x_t):=E(\bfm f_t|\bfm x_t)=\frac{1}{N}\ensuremath{\boldsymbol{\Lambda}}' E(\bfm y_t|\bfm x_t). $$

The normalization ((ref)) above is useful to facilitate the above arguments. In this paper, they are not imposed. Then the same argument shows that $\ensuremath{\boldsymbol{\Lambda}}$ and $\bfm g(\bfm x_t)$ can be identified up to a rotation matrix transformation.

Let

eqnarray*[eqnarray* omitted — 204 chars of source]
assSuppose $\{\bfm f_t,\bfm x_t,\bfm u_t\}_{t\leq T}$ are identically distributed. Assume: (i) Rank condition: $\chi_N>0$. (ii) There are positive constants $\underbar{c}_\Lambda, \bar{c}_\Lambda>0$, so that all the eigenvalues of the $K\times K$ matrix $\bfsym \Sigma_{\Lambda,N}$ are confined in $[\underbar{c}_\Lambda, \bar{c}_\Lambda]$, regardless of whether $N\to\infty$ or not.

Condition (i) is the key condition on the explanatory power of $\bfm x_t$ on factors, where $\chi_N$ represents the “signal strength" of the model. We postpone the discussion of this condition after Theorem (ref). Condition (ii) in Assumption (ref) can be weakened to allow the eigenvalues of $\bfsym \Sigma_{\Lambda, N}$ to slowly decay to zero. While doing so allows some of the factors to be weak, it does not provide any new statistical insights, but would bring unnecessary complications to our results and conditions. Therefore, we maintain the strong version as condition (ii).

Generally, we have the following theorem for identifying $(\ensuremath{\boldsymbol{\Lambda}}, \bfm g(\bfm x_t))$ (up to a rotation transformation).

thmSuppose $E(\bfm u_t|\bfm x_t)=0$, Assumption (ref) holds and $N>K$. Then there is an invertible $K\times K$ matrix $\bfm H$ so that: (i) The columns of $\ensuremath{\boldsymbol{\Lambda}} \bfm H$ are the eigenvectors of $\bfsym \Sigma_{y|x}$ corresponding to the nonzero distinct eigenvalues. (ii) Given $\ensuremath{\boldsymbol{\Lambda}}\bfm H$, $\bfm g(\bfm x_t):=E(\bfm f_t|\bfm x_t) $ satisfies: $$ \bfm H^{-1}\bfm g(\bfm x_t)=[(\ensuremath{\boldsymbol{\Lambda}}\bfm H)'\ensuremath{\boldsymbol{\Lambda}}\bfm H ]^{-1}\ensuremath{\boldsymbol{\Lambda}}\bfm H'E(\bfm y_t|\bfm x_t). $$ (iii) Let $\lambda_K(\bfsym \Sigma_{y|x})$ denote the $K$th largest eigenvalue of $\bfsym \Sigma_{y|x}$, we have $$ \lambda_K(\bfsym \Sigma_{y|x})\geq N\chi_N\underbar{c}_\Lambda. $$ where $\chi_N$ and $\underbar{c}_\Lambda$ are defined in Assumption (ref). In addition, under the normalization conditions that $ E\{E(\bfm f_t|\bfm x_t)E(\bfm f_t|\bfm x_t)'\} $ is a diagonal matrix and that $ \bfsym \Sigma_{\Lambda,N}=\bfm I_K$, we have $ \bfm H=\bfm I_K. $

Discussions of Condition (i) of Assumption (ref)

In the model $$ \bfm f_t=\bfm g(\bfm x_t)+\bfsym \gamma_t,\quad \bfm g(\bfm x_t)=E(\bfm f_t|\bfm x_t), $$ $ \chi_N =\lambda_{\min}(\bfsym \Sigma_{f|x})$ represents the “signal" of the covariate model. We require $\chi_N>0$ so that the rank of $\bfsym \Sigma_{y|x}$ is $K$. Only if this condition holds are we able to identify all the $K$ factor loadings using the eigenvectors corresponding to the nonzero eigenvalues. From the estimation point of view, we are using the PCAs of the estimated $\bfsym \Sigma_{y|x}$, and can only consistently estimate its rank$(\bfsym \Sigma_{y|x})$-number of leading eigenvectors. So this condition is also essential to achieve the consistent estimation of the factor loadings.

Note that requiring $\bfsym \Sigma_{f|x}$ be of full rank might be restrictive in some cases. For instance, consider the linear case: $ E(\bfm f_t|\bfm x_t) =\bfsym \beta\bfm x_t$ for a $K\times d$ coefficient matrix $\bfsym \beta$, also suppose $E\bfm x_t\bfm x_t'$ is of full rank. Then $\bfsym \Sigma_{f|x}=\bfsym \beta E\bfm x_t\bfm x_t' \bfsym \beta'$, and is full-rank only if $d\geq K$. Thus we implicitly require, for linear models, the number of covariates should be at least as many as the number of latent factors. Note that if $E(\bfm f_t|\bfm x_t)$ is highly nonlinear, it is still possible to satisfy the full rank condition even if $d<K$, and we illustrate this in the simulation section. \footnote{Suppose $E(\bfm f_t|\bfm x_t)$ is nonlinear and can be well approximated by a series of orthogonal basis functions $\Phi(\bfm x_t)=(\phi_1(\bfm x_t),...,\phi_J(\bfm x_t))'$, where $E\phi_i(\bfm x_t)\phi_j(\bfm x_t)=1\{i=j\}$, then for some $K\times J$ coefficient $\bfsym \alpha$, we have $E(\bfm f_t|\bfm x_t)\approx \bfsym \alpha' \Phi(\bfm x_t)$ so $E\{E(\bfm f_t|\bfm x_t)E(\bfm f_t|\bfm x_t)'\}\approx \bfsym \alpha\bfsym \alpha'$. For nonlinear functions, it is not stringent to require $\bfsym \alpha\bfsym \alpha'$ be full rank since $K<J$ as $J\to\infty$. }

Definition of the estimators

The above identification strategy motivates us to estimate $\ensuremath{\boldsymbol{\Lambda}}$ and $\bfm g(\bfm x_t)$ respectively by $\widehat{\ensuremath{\boldsymbol{\Lambda}}}$ and $\widehat \bfm g(\bfm x_t)$ as follows. Let $\widehat{\bfsym \Sigma}$ and $\widehat E(\bfm y_t|\bfm x_t)$ be some estimator of $\bfsym \Sigma_{y|x}$ and $E(\bfm y_t|\bfm x_t)$, whose definitions will be clear below. Then the columns of $\frac{1}{\sqrt{N}}\widehat{\ensuremath{\boldsymbol{\Lambda}}}$ are defined as the eigenvectors corresponding to the first $K$ eigenvalues of $\widehat{\bfsym \Sigma}$, and $$\widehat\bfm g(\bfm x_t):=\frac{1}{N}\widehat{\ensuremath{\boldsymbol{\Lambda}}}' \widehat E(\bfm y_t|\bfm x_t). $$ Recall that $\bfm f_t=\bfm g(\bfm x_t)+\bfsym \gamma_t$. We estimate $\bfm f_t$ using least squares: $$ \widehat\bfm f_t:=(\widehat{\ensuremath{\boldsymbol{\Lambda}}}'\widehat{\ensuremath{\boldsymbol{\Lambda}}})^{-1}\widehat{\ensuremath{\boldsymbol{\Lambda}}}'\bfm y_t=\frac{1}{N}\widehat{\ensuremath{\boldsymbol{\Lambda}}}' \bfm y_t. $$ Finally, we estimate $\bfsym \gamma_t$ by: $ \widehat{\bfsym \gamma}_t=\widehat\bfm f_t-\widehat\bfm g(\bfm x_t)=\frac{1}{N}\widehat{\ensuremath{\boldsymbol{\Lambda}}}'(\bfm y_t-\widehat E(\bfm y_t|\bfm x_t)). $ Estimating $\bfm g(\bfm x_t)$ and $\bfsym \gamma_t$ separately allows us to estimate and distinguish the percentage of explained and unexplained components in factors, as well as to quantify the explanatory power of covariates.

Below we introduce the estimators $\widehat{\bfsym \Sigma}$ and $\widehat E(\bfm y_t|\bfm x_t)$ to be used in this paper.

Robust estimation for $\widehat{\bfsym \Sigma}$

Recall that $\bfsym \Sigma_{y|x}=E\{ E(\bfm y_t|\bfm x_t) E(\bfm y_t|\bfm x_t)' \}$, and let us first construct an estimator for $ E(\bfm y_t|\bfm x_t)$ as follows. While many standard nonparametric regressions would work, here we choose an estimator that is robust to the tail-distributions of $\bfm y_t-E(\bfm y_t|\bfm x_t)$.

Let $\Phi(\bfm x_t)=(\phi_1(\bfm x_t),...,\phi_J(\bfm x_t))'$ be a $J\times 1$ dimensional vector of sieve basis. Suppose $E(\bfm y_t|\bfm x_t)$ can be approximated by a sieve representation: $ E(\bfm y_t|\bfm x_t)\approx \bfm B \Phi(\bfm x_t), $ where $\bfm B=(\bfm b_1,...,\bfm b_N)'$ is an $N\times J$ matrix of sieve coefficients. To adapt to different heaviness of the tails of idiosyncratic components, we use the Huber loss function (huber1964robust) to estimate the sieve coefficients. Define $$ \rho(z)=

casesz^2, & |z|<1\\ 2|z|-1, & |z|\geq 1.

$$ For some deterministic sequence $\alpha_T\to\infty$ (adaptive Huber loss), we estimate the sieve coefficients $\bfm B$ by the following convex optimization: $$ \widehat{\bfm b}_i=\arg\min_{b\in\mathbb{R}^J} \frac{1}{T}\sum_{t=1}^T \rho\left( \frac{y_{it}-\Phi(\bfm x_t)'\bfm b}{\alpha_T}\right),\quad \widehat{\bfm b}=(\widehat{\bfm b}_1,...,\widehat{\bfm b}_N)'. $$ We then estimate $\bfsym \Sigma_{y|x}$ by $$ \widehat{\bfsym \Sigma}=\frac{1}{T}\sum_{t=1}^T\widehat E(\bfm y_t|\bfm x_t) \widehat E(\bfm y_t|\bfm x_t)',\quad where \widehat E(\bfm y_t|\bfm x_t)=\widehat{\bfm b}\Phi(\bfm x_t). $$

An alternative method to the robust estimation of $\bfsym \Sigma_{y|x}$ is based on the sieve-least squares, corresponding to the case where $\alpha_T = \infty$. Let $\bfm Y=(\bfm y_1,...,\bfm y_T), $ which is $(N\times T)$, and $$ \bfm P=\Phi'(\Phi\Phi')^{-1}\Phi, (T\times T),\quad \Phi=(\Phi(\bfm x_1),...,\Phi(\bfm x_T)), (J\times T). $$ Then, the sieve least-squares estimator for $\bfsym \Sigma_{y|x}$ is $ \widetilde\bfsym \Sigma= \frac{1}{T} \bfm Y\bfm P\bfm Y'$. While this estimator is attractive due to its closed form, it is not as good as $\widehat\bfsym \Sigma$ when the distribution of $\bfm u_t$ has heavier tails. As expected, our numerical studies in Section (ref) demonstrate that it performs well in light-tailed scenarios, but is less robust to heavy-tailed distributions. Our theories are presented for $\widehat\bfsym \Sigma$, but most of the theoretical findings should carry over to $\widetilde\bfsym \Sigma$.

Choosing $\alpha_T$ and $J$

The selection of the sieve dimension $J$ has been widely studied in the literature, e.g., Li_87, Andrews_91, np_AIC_98, among others. Another tunning parameter is $\alpha_T$, which diverges in order to reduce the biases of estimating the conditional mean when the distribution of $\bfm y_t-E(\bfm y_t|\bfm x_t)$ is asymmetric. Throughout the paper, we shall set

equation[equation omitted — 73 chars of source]

for some constant $C_\alpha>0$, and choose $(J, C_\alpha)$ simultaneously using the multi-fold cross-validation\footnote{One can also allow $\alpha_T$ to depend on $\operatorname{var}(y_{it}|\bfm x_t)$ to allow for different scales across individuals. We describe this choice in the simulation section. In addition, the cross-validation can be based on either in-sample fit for $E(y_{it}|\bfm x_t)$ or out-of-sample forecast, depending on the specific applications. In time series forecasts, one may also consider the time series cross validation TSCV where the training and testing sets are defined through a moving window forecast. }. The specified rate in ((ref)) is due to a theoretical consideration, which leads to the “least biased robust estimation", as we now explain. The Huber-estimator is biased for estimating the mean coefficient in $E(y_{it}|\bfm x_t)$, whose population counterpart is $$ \bfm b_{i,\alpha}:=\arg\min_{\bfm b\in\mathbb{R}^J} E \rho\left( \frac{y_{it}-\Phi(\bfm x_t)'\bfm b}{\alpha_T}\right), $$ As $\alpha_T$ increases, it approaches the limit $\bfm b_i:=\arg\min_{\bfm b\in\mathbb{R}^J} E[y_{it}-\bfm b'\Phi(\bfm x_t)]^2$ with the speed $$ \max_{i\leq N}\|\bfm b_{i,\alpha}-\bfm b_i\|=O(\alpha_T^{-(\zeta_2+1)+\epsilon}) $$ for an arbitrarily small $\epsilon>0$, where $\zeta_2$ is defined in Assumption (ref). Hence the bias decreases as $\alpha_T$ grows. On the other hand, our theory requires the uniform convergence (in $i=1,...,N$) of (for $e_{it}=y_{it}-E(y_{it}|\bfm x_t)$)

equation[equation omitted — 116 chars of source]

where $\dot\rho(\cdot) $ denotes the derivative of $\rho(\cdot)$. It turns out that $\alpha_T$ cannot grow faster than $O(\sqrt{\frac{T}{\log(NJ)}})$ in order to guard for robustness and to have a sharp uniform convergence for ((ref)). Hence the choice ((ref)) leads to the asymptotically least-biased robust estimation.

commentAs for the multi-fold cross validations, we randomly divide the sample into $M$ folds with subsample sizes $\{T_1, \cdots , T_M\}$. Then we can define the in-sample cross-validation criterion for a given pair of $C_\alpha$ and $J$ as $$ CV_{\text{in}}(C_\alpha, J)=\frac{1}{T}\sum\limits_{m=1}^M \sum\limits_{t=1}^{T_m}\sum\limits_{i=1}^N\left(y_{it}-\Phi(\bfm x_t)'\widetilde\bfm b_{i}^{(-m)}(C_\alpha, J)\right)^2, $$ where $\widetilde\bfm b_{i}^{(-m)}(C_\alpha, J)$ is the fitted parameter without using data in $T_m$: $$\widetilde\bfm b_{i}^{(-m)}(C_\alpha, J) =\arg\min_{\bfm b\in\mathbb{R}^J}\frac{1}{T-T_m}\sum_{m'=[M] \setminus m}\sum_{t=1}^{T_{m'}} \rho\left( \frac{y_{it}-\Phi(\bfm x_t)'\bfm b}{C_\alpha\sqrt{\frac{T}{\log(NJ)}} }\right).$$ In practice, let $\cal A$ and $\cal J$ be two sets of gird points for $C_\alpha$ and $J$ respectively. We use the following CV-based $(\widehat{\alpha_T}, \widehat J)$ for $(\alpha_T, J)$: $$ (\widehat{C_\alpha}, \widehat{J})=\arg\min_{(C_\alpha, J) \in {\cal A} \times {\cal J} }CV_{\text{in}}(C_\alpha, J),\quad \widehat{\alpha_T}=\widehat{C_\alpha}\sqrt{\frac{T}{\log(N\widehat J)}} . $$ Alternatively, in out-of-sample forecast applications using estimated factors (see our empirical application). We aim to conduct an $h$-step ahead forecast of a scalar variable $z_{T+h}$ using the data $\{z_t, \bfm x_t, \bfm y_t\}_{t\leq T}$. Let $\widehat z_{t+h|t} (C_\alpha, J)$ be the predicted value of $z_{t+h}$ using the data up to time $t$, which depends on $(C_\alpha, J)$ through the estimated factors. Then we can choose the tuning parameters by minimizing the out-of-sample cross-validation criterion TSCV: $$ CV_{\text{out}}(C_\alpha, J)= \sum\limits_{t=1}^{T-h} \left(z_{t+h}-\widehat z_{t+h|t}(C_\alpha, J)\right)^2. $$

Alternative estimators

Plugging $\bfm f_t=\bfm g(\bfm w_t)+\bfsym \gamma_t$ into ((ref)), we obtain

equation[equation omitted — 204 chars of source]

A closely related model is:

equation[equation omitted — 108 chars of source]

for a nonparametric function $\bfm h(\cdot)$, or simply a linear form $\bfm h(\bfm x_t)=\bfsym \beta\bfm x_t$. Models ((ref)) and ((ref)) were studied in the literature ALS, bai, MW11, where parameters are often estimated using least squares. For instance, we can estimate model ((ref)) by

equation[equation omitted — 200 chars of source]

But this approach is not appropriate in the current context when $\bfm x_t$ almost fully explains $\bfm f_t$ for all $t=1,...,T$. In this case, $\bfsym \gamma_t\approx 0$, and least squares ((ref)) would be inconsistent. \footnote{The inconsistency is due to the fact that $a\ensuremath{\boldsymbol{\Lambda}}\bfsym \gamma_t\approx \ensuremath{\boldsymbol{\Lambda}}\bfsym \gamma_t$ for any scalar $a$ in the case $\bfsym \gamma_t\approx0$. Thus $\ensuremath{\boldsymbol{\Lambda}}$ is not identifiable in the least squares problem.} In addition, $\ensuremath{\boldsymbol{\Lambda}}$ in ((ref)) would be very close to zero because the effects of $\bfm f_t$ would be fully explained by $\bfm h(\bfm w_t)$. As a result, the factors in ((ref)) cannot be consistently estimated onatski2012asymptotics either. We conduct numerical comparisons with this method in the simulation section. In all simulated scenarios, the interactive effect approach gives the worst estimation performance.

Another simpler alternative is to combine $(\bfm x_t, \bfm y_t)$, and apply the classical methods on this enlarged dataset. One potential drawback is that the rates of convergence would not be improved, even if $\bfm x_t$ has strong explanatory power on the factors. Another drawback, as mentioned before, is that $\bfm x_t$ and $\bfm y_t$ can provide very different information (e.g. Fama-French factors versus returns of individual stocks).

comment\subsection{Discussions of other possible alternatives} One of the referees raised a good question that whether we can apply the seemingly more straightforward least squares to estimate the model. Indeed, substitute $\bfm f_t=\bfm g(\bfm x_t)+\bfsym \gamma_t$ to the original model, we reach \begin{equation} \bfm y_t=\ensuremath{\boldsymbol{\Lambda}}\bfm g(\bfm x_t)+\ensuremath{\boldsymbol{\Lambda}}\bfsym \gamma_t+\bfm u_t. \end{equation} Then potentially we could have applied least squares to estimate $\ensuremath{\boldsymbol{\Lambda}}, \bfm g(\cdot), \bfsym \gamma_t$. Alternatively, we could have proposed a model such as \begin{equation} \bfm y_t=\ensuremath{\boldsymbol{\Lambda}}\bfm f_t+ \bfm G(\bfm x_t) +\bfm u_t \end{equation} for some unknown $N\times 1$ dimensional function $\bfm G(\cdot)$, or simply assume a linear form $\bfm G(\bfm x_t)=\bfsym \beta\bfm x_t$ and used least squares: \begin{equation} \min_{\ensuremath{\boldsymbol{\Lambda}}, \bfsym \beta,\{\bfm f_t\}}\sum_{t=1}^T\|\bfm y_t-\ensuremath{\boldsymbol{\Lambda}}\bfm f_t-\bfsym \beta\bfm x_t\|^2. \end{equation} Models ((ref)) and ((ref)) are closely related to the so-called “panel data models with interactive effects" in the literature ALS, bai, MW11, ando2017clustering, and have been often estimated by the least squares like ((ref)). But we would like to emphasize that applying least squares on either ((ref)) or ((ref)) would not be appropriate in the current context. In particular, ((ref)) is in fact not a special case of ((ref)), even if $\bfm g(\cdot)$ is linear. There are two main drawbacks that make model ((ref)), and its associated least squares method, inappropriate in the current context. One is that model ((ref)) requires both $\ensuremath{\boldsymbol{\Lambda}}$ and $\bfm f_t$ be “strong" so that $\ensuremath{\boldsymbol{\Lambda}}'\ensuremath{\boldsymbol{\Lambda}}$ and $\operatorname{cov}(\bfm f_t)$ should not be close to zero. In the current context, however, when $\bfm x_t$ nearly fully explain $\bfm f_t$, we would have $\bfsym \gamma_t\approx 0$, which would violate this requirement, and applying least squares to ((ref)) would be inconsistent. In addition, when $\bfm G(\bfm x_t)$ and $\bfm f_t$ are strongly correlated, $\ensuremath{\boldsymbol{\Lambda}}$ in ((ref)) would be also close to zero. As a result, the factors in ((ref)) cannot be consistently estimated (see onatski2012asymptotics for more discussions). The other drawback is that even if $\ensuremath{\boldsymbol{\Lambda}}$ and $\bfm f_t$ are “strong", and $\bfm G(\bfm x_t)$ in ((ref)) were completely known, we could define $\bfm z_t:= \bfm y_t- \bfm G(\bfm x_t)$, then ((ref)) would become $ \bfm z_t=\ensuremath{\boldsymbol{\Lambda}}\bfm f_t+ \bfm u_t$, the regular factor model without covariates. Therefore, the estimating procedure would be essentially the same as the regular methods and no improvements in any dimension would be achieved. We conduct numerical comparisons with the “interactive effect+least squares" approach in the supplementary document. In all simulated scenarios, this approach yields the worst estimation performance.

Rates of Convergence

commentOne of the popular technical tools in the literature on “spiked covariances” is the perturbation theory (e.g. paul2007asymptotics, birnbaum2013minimax), which provides inequalities that directly upper bound the distance between the eigenvectors and their estimators. But this technique would not provide sharp rates for estimating the low-dimensional functionals of the eigenvectors. For instance, suppose we construct an estimator as a linear transformation of $\widehat{\ensuremath{\boldsymbol{\Lambda}}}$, a typical example is $\widehat\bfm f_t=\frac{1}{N}\widehat{\ensuremath{\boldsymbol{\Lambda}}}'\bfm y_t$, then it can be proved that for some rotation matrix $\bfm H$, and $\ensuremath{\boldsymbol{\Delta}}_\Lambda=\widehat{\ensuremath{\boldsymbol{\Lambda}}}-\ensuremath{\boldsymbol{\Lambda}}\bfm H$, $$ \widehat\bfm f_t-\bfm H^{-1}\bfm f_t=\underbrace{\frac{1}{N}\bfm H'\ensuremath{\boldsymbol{\Lambda}}'\bfm u_t }_{\text{oracle term}}+\underbrace{\frac{1}{N}\ensuremath{\boldsymbol{\Delta}}_\Lambda'\bfm u_t}_{\text{effect of estimating $\ensuremath{\boldsymbol{\Lambda}}$}} $$ where the oracle term is the statistical error if $\ensuremath{\boldsymbol{\Lambda}}$ were known. The perturbation theory however, would only bound $\frac{1}{N}\ensuremath{\boldsymbol{\Delta}}_\Lambda'\bfm u_t$ by $\|\frac{1}{N}\ensuremath{\boldsymbol{\Delta}}_\Lambda\|\| \bfm u_t\|$ which is not sufficiently sharp, and may even obtain a rate dominating the oracle term. In fact, almost the entire literature only focuses on the eigenvectors $\ensuremath{\boldsymbol{\Lambda}}$ themselves, rather than their low-dimensional functionals. Instead, we use a different approach by deriving a Bahadur expansion bahadur1966note of the estimated eigenvectors for the spiked matrices, in the following form: \begin{eqnarray} \widehat{\ensuremath{\boldsymbol{\Lambda}}}-\ensuremath{\boldsymbol{\Lambda}}\bfm H &=&\frac{1}{NT}\sum_{t=1}^T\ensuremath{\boldsymbol{\Lambda}} \bfm g(\bfm x_t)\Phi(\bfm x_t)'\bfm A\sum_{i=1}^N\frac{1}{T}\sum_{s=1}^T\Phi(\bfm x_s)'\dot{\rho}(\alpha_T^{-1}e_{is})\alpha_T \widehat{\ensuremath{\boldsymbol{\Lambda}}} \widetilde\bfm V^{-1} + \widetilde{\ensuremath{\boldsymbol{\Delta}}} \end{eqnarray} where $\dot{\rho}$ denotes the derivative of the Huber's loss function: $$ \dot{\rho}(z)=\begin{cases} 2z, & |z|<1\\ 2\operatorname{sgn}(z), & |z|\geq 1. \end{cases} $$ Here $\operatorname{sgn}(z)$ denotes the sign function; $\bfm e_t:=\bfm y_t-E(\bfm y_t|\bfm x_t)=(e_{1t},...,e_{Nt})'$; $\bfm A $ is the Hessian matrix of the expected Huber's loss function; $\widetilde\bfm V$ is a $K$-dimensional diagonal matrix of the eigenvalues of $\widehat{\bfsym \Sigma}/N$. The second term $ \widetilde{\ensuremath{\boldsymbol{\Delta}}} $ is a higher order random term. Such an expansion allows us to derive a much sharper bound for low-dimensional functionals of $\ensuremath{\boldsymbol{\Delta}}_\Lambda$ such as $\frac{1}{N}\ensuremath{\boldsymbol{\Delta}}_\Lambda'\bfm u_t$. It is also potentially useful for deriving the limiting distributions of the estimators, which cannot be achieved by the perturbation theory. \footnote{While deriving the limiting distributions for the estimated factors and loadings is out of the scope of this paper, we expect that they can be derived based on the Bahadur expansion, and shall leave it for the future research.} Throughout the paper, we assume $T$ grows to infinity, while $K=\dim(\bfm f_t)$ and $d=\dim(\bfm x_t)$ are constant. In addition, $N$ may either grow or stay constant.

Assumptions

Let $ e_{it}:=y_{it}-E(y_{it}|\bfm x_t). $ Suppose the conditional distribution of $e_{it}$ given $\bfm x_t=\bfm x$ is absolutely continuous for almost all $\bfm x$, with a conditional density $g_{e, i}(\cdot|\bfm x)$.

ass[Tail distributions] (i) There are $\zeta_1, \zeta_2>2$, $C>0$ and $M>0$, so that for all $x>M$, \begin{equation} \sup_{\bfm x}\max_{i\leq N} g_{e,i}(x|\bfm x)\leq Cx^{-\zeta_1},\quad \sup_{\bfm x}\max_{i\leq N} E(e_{it}^21\{|e_{it}|>x\}|\bfm x_t=\bfm x)\leq Cx^{-\zeta_2}. \end{equation} (ii) $\Phi(\bfm x_t)$ is a sub-Gaussian vector, that is, there is $L>0$, for any $\ensuremath{\boldsymbol{\nu}}\in\mathbb{R}^J$ so that $\|\ensuremath{\boldsymbol{\nu}}\|=1$, $$ P(|\ensuremath{\boldsymbol{\nu}}'\Phi(\bfm x_t)|>x)\leq \exp(1-x^2/L),\quad\forall x\geq0. $$
ass[Sieve approximations] (i) For $k=1,...,K$, let $\bfm v_k=\arg\min_{\bfm v}E(f_{kt}-\bfm v'\Phi(\bfm x_t))^2$. Then there is $\eta\geq 2$, as $J\to\infty$, $$ \max_{k\leq K}\sup_{\bfm x}|E(f_{tk}|\bfm x_t=\bfm x)-\bfm v_k'\Phi(\bfm x)|=O(J^{-\eta}). $$ (ii) There are $c_1, c_2>0$ so that \begin{eqnarray*} &&c_1\leq\lambda_{\min}(E\Phi(\bfm x_t)\Phi(\bfm x_t)')\leq\lambda_{\max}(E\Phi(\bfm x_t)\Phi(\bfm x_t)')\leq c_2. \end{eqnarray*}

Recall $\bfsym \gamma_t=\bfm f_t-E(\bfm f_t|\bfm x_t)$. Let $\gamma_{kt}$ be its $k$ th component.

ass[Weak dependences] (i) (serial independence) $\{\bfm f_t, \bfm u_t, \bfm x_t\}_{t\leq T}$ is independent and identically distributed; (ii) (weak cross-sectional dependence) For some $C>0$, $$\sup_{\bfm x, \bfm f} \max_{i\leq N}\sum_{j=1}^N|E(u_{it}u_{jt}|\bfm x_t=\bfm x, \bfm f_t=\bfm f)| <C.$$ (iii) $ E(\bfm u_t|\bfm f_t, \bfm x_t)=0$, $\max_{i\leq N}\|\ensuremath{\boldsymbol{\lambda}}_i\|<C$, and $\operatorname{cov}(\bfsym \gamma_t|\bfm x_t)=\operatorname{cov}(\bfsym \gamma_t)$ almost surely, where $\operatorname{cov}(\bfsym \gamma_t|\bfm x_t)$ denotes the conditional covariance matrix of $\bfsym \gamma_t$ given $\bfm x_t$, assumed to exist.

Recall that $$\bfsym \Sigma_{f|x}:=E\{E(\bfm f_t|\bfm x_t)E(\bfm f_t|\bfm x_t)'\} ,\quad \chi_N:=\lambda_{\min}( \bfsym \Sigma_{f|x} ) .$$

ass[Signal-noise] (i) There is $C>0$, $$ \frac{\lambda_{\max}( \bfsym \Sigma_{f|x})}{\lambda_{\min}( \bfsym \Sigma_{f|x} )}<C, \quad \frac{\lambda_{\max}( E\{\Phi(\bfm x_t)E(\bfm f_t|\bfm x_t)'E(\bfm f_t|\bfm x_t)\Phi(\bfm x_t)'\})}{\lambda_{\min}(\bfsym \Sigma_{f|x} )}<C. $$ (ii) There is $v>1$, so that $\max_{k\leq K}E[E(\gamma_{kt}^4|\bfm x_t)]^v<\infty$.\\ (iii) We have $J^3\log ^2N=O(T)$ and $$ J^{2}/T+J^{-\eta}+\sqrt{(\log N)/T}\ll \chi_N. $$

Assumption (ref) allows distributions with relatively heavy tails on $y_{it}-E(y_{it}|\bfm x_t)$. We still require sub-Gaussian tails for the sieve basis functions. Assumption (ref) is regarding the accuracy of sieve approximations for nonparametric functions. Assumption (ref) strengthens Assumption (ref). We respectively regard $ \lambda_{\min}(\bfsym \Sigma_{f|x})$ and $\operatorname{cov}(\bfsym \gamma_t)$ as the “signal" and “noise" when using $\bfm x_t$ to explain common factors. The explanatory power is measured by these two quantities.

Assumption (ref) (i) requires serial independence, and we admit that it can be restrictive in applications. Allowing for serial dependence is technically difficult due to the non-smooth Huber's loss. To obtan the Bahadur representation of the estimated eigenvectors, we rely on the symmetrization and contraction theorems (e.g., VW), which requires the data be independently distributed. Nevertheless, the idea of using covariates would still be applicable for serial dependent data. For instance, it is not difficult to allow for weak serial correlations when the data are not heavy-tailed, by using the sieve least squares estimator $\widetilde\bfsym \Sigma$ (introduced in Section (ref)) in place of the Huber's estimator $\widehat{\bfsym \Sigma}$. We conduct numerical studies when the data are serially correlated in the simulations, and find that the proposed methods continue to perform well in the presence of serial correlations.

Rates of convergence

We present the rates of convergence in the following theorems, and discuss the statistical insights in the next subsection. Recall $\widehat{\ensuremath{\boldsymbol{\Lambda}}}=(\widehat{\ensuremath{\boldsymbol{\lambda}}}_i: i\leq N)$.

thm[Loadings] Under Assumptions (ref)--(ref), there is an invertible matrix $\bfm H$, as $ T, J\to\infty$, and $N$ either grows or stays constant, \begin{eqnarray} \frac{1}{N}\sum_{i=1}^N\|\widehat{\ensuremath{\boldsymbol{\lambda}}}_i-\bfm H'\ensuremath{\boldsymbol{\lambda}}_i\|^2&=&O_P\left(\frac{J}{T}+\frac{1}{J^{2\eta-1}}\right)\chi_N^{-1},\\ \max_{i\leq N}\|\widehat{\ensuremath{\boldsymbol{\lambda}}}_i-\bfm H'\ensuremath{\boldsymbol{\lambda}}_i\|&=&O_P\left( \sqrt{\frac{J\log N}{T}}+\frac{1}{J^{\eta-1/2}}\right)\chi_N^{-1/2}. \cr \end{eqnarray}

The optimal rate for $J$ in ((ref)) is $J\asymp T^{1/(2\eta)}$, which results in

equation[equation omitted — 188 chars of source]

Here $\eta$ represents the smoothness of $E(\bfm f_t|\bfm x_t=\cdot)$, as defined in Assumption (ref).

Define $$ J^{*}= \min\left\{ (TN)^{1/(2\eta)}, ( \frac{T}{\log N})^{1/(1+\eta)} \right\}. $$

thm[Factors] Let $J\asymp J^{*}$. Suppose $(J^*)^3\log^2N=O(T)$, and Assumptions (ref)--(ref) hold. For $\bfm H$ in Theorem (ref), as $ T\to\infty$, and $N$ either grows or stays constant, we have \begin{equation*} \frac{1}{T}\sum_{t=1}^T\|\widehat\bfm g(\bfm x_t)-\bfm H^{-1}\bfm g(\bfm x_t)\|^2= O_P\left( r_{T, N}^* + (\frac{\log N}{T})^{2-\frac{3}{1+\eta}} \right), \end{equation*} where $r_{T, N}^* = \frac{J^{*2}}{T^2}\chi_N^{-1} + {\frac{J^*\|\operatorname{cov}(\bfsym \gamma_t)\|}{T}}+ (\frac{1}{TN})^{1-\frac{1}{2\eta}}$ and \begin{eqnarray} \frac{1}{T}\sum_{t=1}^T\| \widehat{\bfsym \gamma}_t-\bfm H^{-1}\bfsym \gamma_t \|^2&=&O_P\left( r_{T, N}^* + (\frac{\log N}{T})^{2-\frac{4}{1+\eta}} \right)\chi_N^{-1}\cr &&+O_P\left(\frac{1}{N}\right) . \end{eqnarray}

These two convergences imply the rate of convergence of the estimated factors due to $\widehat\bfm f_t=\widehat \bfm g(\bfm x_t)+\widehat{\bfsym \gamma}_t$.

remFor a general $J$, the rates of convergence of the two factor components are \begin{equation} \frac{1}{T}\sum_{t=1}^T\|\widehat\bfm g(\bfm x_t)-\bfm H^{-1}\bfm g(\bfm x_t)\|^2= O_P\left( r_{T, N} +\frac{J^3\log^2 N}{T^2} \right), \end{equation} where $r_{T, N} = \frac{J^2}{T^2}\chi_N^{-1} +\frac{J\| \operatorname{cov}(\bfsym \gamma_s)\|}{T}+J^{1-2\eta}+\frac{J}{TN}$ and \begin{eqnarray} \qquad \qquad \frac{1}{T}\sum_{t=1}^T\| \widehat{\bfsym \gamma}_t-\bfm H^{-1}\bfsym \gamma_t \|^2 = O_P\left( r_{T, N} + \frac{ J^4\log^2 N }{T^2} \right) \chi_N^{-1} + O_P\left(\frac{1}{N}\right). \end{eqnarray} In fact $J\asymp J^*$ is the optimal choice in ((ref)) ignoring the terms involving $\|\operatorname{cov}(\bfsym \gamma_s)\|$ and $\chi_N$. The convergence rates presented in Theorem (ref) are obtained from ((ref)) and ((ref)) with this choice of $J$. The presented rates connect well with the literature on both standard nonparametric sieve estimations and the high-dimensional factor models. To illustrate this, we discuss in more detail about the rate of convergence in ((ref)). This rate is given by: $$ O_P\left( \underbrace{\frac{J^2}{T^2}\chi_N^{-1}}_{\substack{\text{effect of} \\ \text{estimating $\ensuremath{\boldsymbol{\Lambda}}$}}} +\underbrace{\frac{J\| \operatorname{cov}(\bfsym \gamma_s)\|}{T}+\frac{J}{TN}+J^{1-2\eta}}_{\substack{\text{nonparametric sieve}\\ \text{estimation error}}} +\underbrace{\frac{J^3\log^2 N}{T^2}}_{\substack{\text{higher order from} \\ \text{Huber's M-estimation}}} \right). $$ More specifically, we have, for $\bfm e_t =\ensuremath{\boldsymbol{\Lambda}}\bfsym \gamma_t+\bfm u_t$, \begin{equation} \bfm y_t=\ensuremath{\boldsymbol{\Lambda}}\bfm g(\bfm x_t) +\bfm e_t,\quad E(\bfm e_t|\bfm x_t)=0. \end{equation} If $\ensuremath{\boldsymbol{\Lambda}}$ were known, we would estimate $\bfm g(\cdot)$ by regressing the estimated $E(\bfm y_t|\bfm x_t)$ on $ \ensuremath{\boldsymbol{\Lambda}}$. Then standard nonparametric results show that the rate of convergence in this “oracle sieve regression” (knowing $\ensuremath{\boldsymbol{\Lambda}}$) would be $$ \frac{J\| \operatorname{cov}(\bfsym \gamma_s)\|}{T}+\frac{J}{TN} +J^{1-2\eta}. $$ As we do not observe $\ensuremath{\boldsymbol{\Lambda}}$, we are running the regression ((ref)) with $\widehat{\ensuremath{\boldsymbol{\Lambda}}}$ in place of $\ensuremath{\boldsymbol{\Lambda}}$. This leads to an additional term $\frac{J^2}{T^2}\chi_N^{-1}$ representing the effect of estimating $\ensuremath{\boldsymbol{\Lambda}}$, which also depends on the strength of the signal $\chi_N$. Finally, Huber's M-estimation to estimate $E(\bfm y_t|\bfm x_t)$ gives rise to a higher order term $\frac{J^3\log^2 N}{T^2}$, and is often negligible.

The signal-noise regimes

We see that the rates depend on $\operatorname{cov}(\bfsym \gamma_t)$ and $\chi_N$. Because $E\bfm f_t\bfm f_t'=\bfsym \Sigma_{f|x}+\operatorname{cov}(\bfsym \gamma_t)$, they are related through

equation[equation omitted — 89 chars of source]

for some $c, C_1 >0$, assuming that there is $c>0$ so that $\|E\bfm f_t\bfm f_t'\|>c$. For comparison, we state the rates of convergence of the benchmark PCA estimators: (e.g., SW02, bai03) there is a rotation matrix $\tilde\bfm H$, so that the PCA estimators $( \widetilde\ensuremath{\boldsymbol{\lambda}}_i, \widetilde\bfm f_t)$ satisfy:

equation[equation omitted — 309 chars of source]

The first interesting phenomena we observe is that both the estimated loadings and $\bfm g(\bfm x_t)$ are consistent even if $N$ is finite, due to the “exact identification". In contrast, the PCA estimators requires a growing $N$. For more detailed comparisons, we consider three regimes based on the explanatory power of the factors using $\bfm x_t$. To simplify our discussions, we consider the rate-optimal choices of $J$, and ignore the sieve approximation errors, so $\eta$ is treated sufficiently large.

Regime I: strong explanatory power: $\|\operatorname{cov}(\bfsym \gamma_t)\|\to 0.$ Because of ((ref)), $\chi_N$ is bounded away from zero. In this case, ((ref))-((ref)) approximately imply (for sufficiently large $\eta$):

eqnarray*[eqnarray* omitted — 558 chars of source]

Compared to the rates of the usual PCA estimators in ((ref)), either the new estimated loadings (when $N=o(T)$) or the new estimated factors (when $T=o(N)$) have a faster rate of convergence. Moreover, if $\|\operatorname{cov}(\bfsym \gamma_t)\|= o((TN)^{-1}+T^{-2}\log^2N)$, then $\widehat\bfm g(\bfm x_t)$ directly estimates the latent factor at a very fast rate of convergence: $$ \frac{1}{T}\sum_{t=1}^T\|\widehat\bfm g(\bfm x_t)-\bfm H^{-1}\bfm f_t\|^2=O_P\left( \frac{1}{TN}+ (\frac{\log N}{T})^{2 } \right). $$ The improved rates are reasonable due to the strong explanatory powers from the covariates.

Regime II: mild explanatory power: $\|\operatorname{cov}(\bfsym \gamma_t)\| $ is bounded away from zero; $\chi_N$ is either bounded away from zero or decays slower than $ \frac{N}{T}$ in the case $N=o(T)$. In this regime, $\bfm x_t$ partially explains the factors, yet the unexplainable components are not negligible. ((ref))-((ref)) approximately become:

eqnarray[eqnarray omitted — 319 chars of source]

We see that the rate for the estimated loadings is still faster than the PCA when $N$ is relatively small compared to $T$, while the rates for the estimated factors are the same. This is because $$ \underbrace{\frac{1}{T}\chi_N^{-1}}_{\text{new rate for loadings}}\ll \underbrace{\frac{1}{N}}_{\text{ PCA rate for loadings}} $$ $$ \underbrace{\frac{1}{T}\chi_N^{-1}+ \frac{1}{N}}_{\text{new rate for factors}}\asymp \underbrace{\frac{1}{N}}_{\text{ PCA rate for factors}}. $$ On one hand, due to the explanatory power from the covariates, the loadings can be estimated well without having to consistently estimate the factors. On the other hand, as the covariates only partially explain the factors, we cannot improve rates of convergence in estimating the unexplainable components in the latent factors. However, since $\bfsym \gamma_t$ has smaller variability than $\bfm f_t$, it can still be better estimated in terms of a smaller constant factor.

Regime III: weak explanatory power: $\chi_N\to 0$ and decays faster than $ \frac{N}{T}$ when $N\ll T$. In this case, we have

eqnarray*[eqnarray* omitted — 233 chars of source]

While the new estimators are still consistent, they perform worse than PCA. This finding is still reasonable because the signal is so weak that the conditional expectation $E(\bfm y_t|\bfm x_t)$ loses useful information of the factors/loadings. Consequently, estimation efficiency is lost when running PCA on the estimated covariance $ E\{ E(\bfm y_t|\bfm x_t) E(\bfm y_t|\bfm x_t)' \}$.

In summary, improved rates of convergence can be achieved so long as the covariates can (partially) explain the latent factors, this corresponds to either the mild or the strong explanatory power case. The degree of improvements depend on the strength of the signals. In particular, the consistent estimation for factor loadings can also be achieved even under finite $N$. On the other hand, when the explanatory power is too weak, the rates of convergence would be slower than those of the benchmark estimator.

Testing the Explanatory Power of Covariates

We aim to test: (recall that $\bfsym \gamma_t=\bfm f_t-E(\bfm f_t|\bfm x_t)$)

equation[equation omitted — 74 chars of source]

Under $H_0$, $\bfm f_t=E(\bfm f_t|\bfm x_t)$ over the entire sampling period $t=1,..., T$, implying that observed covariates $\bfm x_t$ fully explain the true factors $\bfm f_t$. In empirical applications with “observed factors", what have been often used are in fact $\bfm x_t$. Hence our proposed test can be applied to empirically validate the explanatory power of these “observed factors".

The Fama-French three-factor model FF is one of the most celebrated ones in empirical asset pricing. They modeled the excess return $r_{it}$ on security or portfolio $i$ for period $t$ as $$ r_{it} =\alpha_i +b_i r_{Mt} + s_i \mbox{SMB}_t +h_i \mbox{HML}_t + u_{it}, $$ where $r_{Mt}, \mbox{SMB}_t$ and $\mbox{HML}_t$ respectively represent the the excess returns of the market, the difference of returns between stocks with small and big market capitalizations (“small minus big"), and the difference of returns between stocks with high book to equity ratios and those with low book to equity ratios (“high minus low"). Ever since its proposal, there is much evidence that the three-factor model can leave the cross-section of expected stock returns unexplained. Different factor definitions have been explored, e.g., carhart1997persistence and novy2013other. fama2015five added profitability and investment factors to the three-factor model. They conducted GRS tests GRS on the five-factor models and its different variations. Their tests “reject all models as a complete description of expected returns".

On the other hand, the Fama-French factors, though imperfect, are good proxies for the true unknown factors. Consequently, they form a natural choice for $\bfm x_t$. These observables are actually diversified portfolios, which have explanatory power on the latent factors $\bfm f_t$, as supported by financial economic theories as well as empirical studies. The test proposed in this validates the specification of these common covariates as “factors".

The Test Statistic

Our test is based on a Wald-type weighted quadratic statistic $$ S(\bfm W):=\frac{N}{T}\sum_{t=1}^T\widehat{\bfsym \gamma}_t'\bfm W\widehat{\bfsym \gamma}_t=\frac{1}{TN}\sum_{t=1}^T(\bfm y_t-\widehat E(\bfm y_t|\bfm x_t))'\widehat{\ensuremath{\boldsymbol{\Lambda}}}\bfm W\widehat{\ensuremath{\boldsymbol{\Lambda}}}'(\bfm y_t-\widehat E(\bfm y_t|\bfm x_t)). $$ The weight matrix normalizes the test statistic, taken as $\bfm W=$ AVar$(\sqrt{N}\widehat{\bfsym \gamma}_t)^{-1}$, where AVar$(\widehat{\bfsym \gamma}_t)$ represents the asymptotic covariance matrix of $\widehat{\bfsym \gamma}_t$ under the null, and is given by $$ \text{AVar}(\sqrt{N}\widehat{\bfsym \gamma}_t)=\frac{1}{N}\bfm H' \ensuremath{\boldsymbol{\Lambda}}'\bfsym \Sigma_u\ensuremath{\boldsymbol{\Lambda}}\bfm H. $$ As $\bfsym \Sigma_u$ is a high-dimensional covariance matrix, to simplify the technical arguments, in this section we assume $\{u_{it}\}$ to be cross-sectionally uncorrelated, and estimate $\bfsym \Sigma_u$ by: $$ \widehat{\bfsym \Sigma}_u=\operatorname{diag}\{\frac{1}{T}\sum_{t=1}^T\widehat u_{it}^2, i=1,..., N\}, \quad \widehat u_{it}=y_{it}-\widehat{\ensuremath{\boldsymbol{\lambda}}}_i'\widehat\bfm f_t. $$ The feasible test statistic is defined as $$ S := S(\widehat\bfm W),\quad \widehat\bfm W:=(\frac{1}{N}\widehat{\ensuremath{\boldsymbol{\Lambda}}}'\widehat{\bfsym \Sigma}_u\widehat{\ensuremath{\boldsymbol{\Lambda}}})^{-1}. $$ We reject the null hypothesis for large values of $S$. It is straightforward to allow $\bfsym \Sigma_u$ to be a non-diagonal but a sparse covariance, and proceed as in Bickel08a. We expect the asymptotic analysis to be quite involved, and do not pursue it in this paper.

We show that the test statistic has the following asymptotic expansion: $$ S= \bar S+ o_P(\frac{1}{\sqrt{T}}), $$ where $$ \bar S=\frac{1}{T}\sum_{t=1}^T \bfm u_t'\ensuremath{\boldsymbol{\Lambda}} (\ensuremath{\boldsymbol{\Lambda}}'\bfsym \Sigma_u\ensuremath{\boldsymbol{\Lambda}})^{-1}\ensuremath{\boldsymbol{\Lambda}}'\bfm u_t. $$ Thus the limiting distribution is determined by that of $\bar S$. Note that a cross-sectional central limit theorem implies, as $N\to\infty$, $$ (\frac{1}{N}\ensuremath{\boldsymbol{\Lambda}}'\bfsym \Sigma_u\ensuremath{\boldsymbol{\Lambda}})^{-1/2}\frac{1}{\sqrt{N}}\bfm u_t'\ensuremath{\boldsymbol{\Lambda}}\to^d\mathcal{N}(0,\bfm I_K). $$ Hence each component of $\bar S$ can be roughly understood as $\chi^2$-distributed with degrees of freedom $K$ being the number of common factors, whose variance is $2K$. This motivates the following assumption.

assSuppose $ \frac{ 1}{T}\sum_{t=1}^T \operatorname{var}(\bfm u_t'\ensuremath{\boldsymbol{\Lambda}} (\ensuremath{\boldsymbol{\Lambda}}'\bfsym \Sigma_u\ensuremath{\boldsymbol{\Lambda}})^{-1}\ensuremath{\boldsymbol{\Lambda}}'\bfm u_t )\to2K$ as $T, N\to\infty$.

We now state the null distribution in the following theorem.

thmSuppose $\{u_{it}\}_{i\leq N}$ is cross-sectionally independent, and Assumption (ref) and assumptions of Theorem (ref) hold. Then, when $J^4N\log N =o(T^{3/2})$, $T=o(N^2)$, $N\sqrt{T}=o(J^{2\eta-1})$, as $T, N\to\infty$, $$\sqrt{\frac{T}{2K}} (S-K)\to^d\mathcal{N}(0,1). $$

Testing market risk factors for S&P 500 returns

\numberwithin{table}{section} \numberwithin{figure}{section}

We test the explanatory power of the observable proxies for the true factors using S&P 500 returns. For each given group of observable proxies, we set the number of common factors $K$ equals the number of observable proxies. We calculate the excess returns for the stocks in S&P 500 index that are collected from CRSP. We consider three groups of proxy factors ($\bfm x_t$) with increasing information: (1) Fama-French 3 factors (FF3); (2) Fama-French 5 factors (FF5); and (3) Fama-French 5 factors plus 9 sector SPDR ETF's (FF5+ETF9). Here the sector SPDR ETF's, which are intended to track the 9 largest S&P sectors. The detailed descriptions of sector SPDR ETF's are listed in Table (ref).

table[table omitted — 479 chars of source]
figure[figure omitted — 339 chars of source]
figure[figure omitted — 327 chars of source]

We consider tests using both daily and monthly data. For the daily data, we collect 393 stocks that have complete daily closing prices from January 2005 to December 2013, with a time span of 2265 trading days. We apply moving window tests with the window size ($T$) equals one month, three months or six months. The testing window moves one trading day forward per test. Within each testing window, we calculate the standardized test statistic $S$ for three groups of proxy factors.

As for the monthly excess returns, we use stocks that have complete record from January 1980 to December 2012, which contains 202 stocks with a time span of 396 months. Here we only consider the first two groups of proxy factors as sector SPDR ETF's are introduced since 1998. The window size equals sixty months and moves one month forward per test. Within each testing window, besides standardized test statistic and p-value, we also estimate the volatility of $\bfsym \gamma_t$, the part of factors that can not be explained by $\bfm x_t$ as: $$ \widehat{\mbox{Vol}}(\bfsym \gamma_t)=\frac{1}{21T}\sum\limits_{t=1}^T\widehat\bfsym \gamma_t'\widehat\bfsym \gamma_t, $$ where there are 21 trading days per month. The sieve basis is chosen as the additive Fourier basis with $J=5$. We set the tuning parameter $\alpha_T=C\sqrt{\frac{T}{\log(NJ)}}$ with constant $C$ selected by the 5-fold cross validation.

For the daily data, the plots of $S$ under various scenarios are reported in Figure (ref). Under all scenarios, the null hypothesis ($H_0: \operatorname{cov}(\bfsym \gamma_t)=0$) is rejected as $S $ is always larger than the critical value 1.96. This suggests a strong evidence that the proxy factors can not fully explain the estimated common factors. Under all window sizes, a larger group of proxy factors tends to yield smaller statistics, demonstrating stronger explanatory power for estimated common factors. Also, we find the test statistics increase while the window size increases.

The results for the monthly data are reported in Figure (ref). For both Fama-French 3 factors and 5 factors, the null hypothesis is rejected most of the time except in early 1980s and 1990s. When the null hypothesis is accepted, Fama-French 5 factors tend to yield larger p-values. The estimated volatility of unexplained part are close to zero over these two periods. For the rest of the time, the standardized test statistics are much larger than the critical value 1.96 and hence the p-values are close to zero. Also the estimated volatilities are not close to zero. This indicates the proxy factors can not fully explain the estimated common factors during these testing periods.

Forecast the excess return of US government bonds

We apply our method to forecast the excess return of U.S. government bonds. The bond excess return is the one-year bond return in excess of the risk-free rate. To be more specific, we buy an $n$ year bond, sell it as an $n-1$ year bond in the next year and excess the one-year bond yield as the risk-free rate. Let $p_t^{(n)}$ be the log price of an $n$-year discount bond at time $t$. Denote $\zeta_t^{(n)} \equiv -\frac{1}{n}p_t^{(n)}$ as the log yield with $n$ year maturity, and $r_{t+1}^{(n)} \equiv p_{t+1}^{(n-1)}-p_t^{(n)}$ as the log holding period return. The goal of one-step-ahead forecast is to forecast $z_{T+1}^{(n)}$, the excess return with maturity of $n$ years in period $T+1$, where $$ z_{t+1}^{(n)}=r_{t+1}^{(n)}-\zeta_t^{(1)},\quad t=1, \ \cdots \ , T.$$

For a long time, the literature has found a significant predictive power of the excess returns of U.S. government bonds. For instance, ludvigson2009macro, ludvigson2010factor predicted the bond excess returns with observable variables based on a factor model using 131 (disaggregated) macroeconomics variables. They achieved the out-of-sample $R^2 \approx 21\%$ when forecasting one year excess bond return with maturity of two years. Using the proposed method, this section develops a new way of incorporating the explanatory power of the observed characteristics, and investigates the robustness of the conclusions in existing literature.

We analyze monthly data spanned from January 1964 to December 2003, which is available from the Center for Research in Securities Prices (CRSP). The factors are estimated from a macroeconomic dataset consisting of 131 disaggregated macroeconomic time series ludvigson2010factor. The covariates $\bfm x_t$ are 8 aggregated macro-economic time series, listed in Table (ref).

table[table omitted — 585 chars of source]

Heavy-tailed data and robust estimations

We first examine the excess kurtosis for the time series to assess the tail distributions. The left panel of Figure (ref) shows 43 among the 131 series have excess kurtosis greater than 6. This indicates the tails of their distributions are fatter than the $t$-distribution with degrees of freedom 5. On the other hand, the right panel of Figure (ref) reports the histograms of excess kurtosis of the “fitted data" $\widehat{E}(\bfm y_t|\bfm x_t)$ (the robust estimator of $E(\bfm y_t|\bfm x_t)$ using Huber loss), which demonstrates that most series in the fitted data are no longer severely heavy-tailed.

The tuning parameter in the Huber loss is of order $\alpha_{T}=C_\alpha \sqrt{\frac{T}{\log(NT)}}$. In this study, the constant $C_{\alpha}$ and the degree of sieve approximation $J$ are selected by the out-of-sample 5-fold cross validation as described in Section (ref).

figure[figure omitted — 366 chars of source]

Forecast results

We denote our proposed method by SPCA (smoothed PCA), and compare it with SPCA-LS (which uses $\widetilde\bfsym \Sigma$, the least-squares based smoothed PCA, described in Section (ref)) and the benchmark PCA. We conduct one-month-ahead out-of-sample forecast of the bond excess returns. The forecast uses the information in the past 240 months, starting from January 1984 and rolling forward to December 2003. We compare three approaches to estimating the factors: SPCA, SPCA-LS, and the usual PCA. Also we consider two forecast models as follows:

eqnarray[eqnarray omitted — 317 chars of source]

where $\alpha$ is the intercept and $h$ is a nonparametric function. The covariate $\bfm W_t$ is either $\bfm f_t$ or an augmented vector $(\bfm f_t',\bfm x_t')'$. Here, the latent factors $\bfm f_t$ are used by the three methods mentioned above in order to compare their effectiveness. The multi-index model allows more general nonlinear forecasts and are estimated by using the sliced inverse regression li1991sliced. The number of indices $L$ is estimated by the ratio-based method suggested in LamYao and is usually 2 or 3. We approximate $h$ using a weighted additive model $h({\ensuremath{\boldsymbol{\psi}}}_1' \bfm W_t,\ \cdots , \ {\ensuremath{\boldsymbol{\psi}}}_L' \bfm W_t)=\sum_{l=1}^L g_l({\ensuremath{\boldsymbol{\psi}}}_l' \bfm W_t)$. Each individual nonparametric function $g_l(\cdot)$ is smoothed by the local linear approximation.

The performance of each method is assessed by the out-of-sample $R^2$. Let $\widehat z_{T+t+1|T+t}$ be the forecast of $z_{T+t+1}$ using the data of the previous $T$ months: $1+t,...,T+t$ for $T=240$ and $t=0,..., 239$. The forecast performance is assessed by the out-of-sample $R^2$, defined as $$ R^2=1- \frac{\sum\limits_{t=0}^{239}(z_{T+t+1}-\widehat{z}_{T+t+1|T+t})^2}{\sum\limits_{t=0}^{239}(z_{T+t+1}-\bar{z}_t)^2}, $$ where $\bar{z}_t$ is the sample mean of $z_t$ over the sample period $[1+t, T+t]$. The $R^2$ of various methods are reported in Table (ref). We notice that factors estimated by SPCA and SPCA-LS can explain more variations in bond excess returns with all maturities than the ones estimated by PCA. SPCA yields a 44.6% out-of-sample $R^2$ for forecasting the bond excess returns with two year maturity, which is much higher than the best out-of-sample predictor found in ludvigson2009macro. It is also observed that the forecast based on either SPCA or SPCA-LS cannot be improved by adding any covariate in $\bfm x_t$. We argue that, in this application, the information of $\bfm x_t$ should be mainly used as the explanatory power for the factors.

We summarize the observed results in the following aspects:

enumerate• The factors estimated using additional covariates lead to significantly improved out-of-sample forecast on the US bond excess returns compared to the ones estimated by PCA. • As many series in the panel data are heavy-tailed, the robust-version of our method (SPCA) can result in improved out-of-sample forecasts. • The multi-index models yield significantly larger out-of-sample $R^2$'s than those of the linear forecast models. • The observed covariates $\bfm x_t$ (e.g. forward rates, employment and inflation) contain strong explanatory powers of the latent factors. The gain of forecasting bond excess returns is more substantial when these covariates are incorporated to estimate the common factors (using the proposed procedure) than directly used for forecasts.
table[table omitted — 1,608 chars of source]
comment\begin{table}[htbp] \begin{center} \caption{Forecast out-of-sample $R^2$ (%) for multi-index model: the larger the better. } \begin{tabular}{c|cccc|cccc|cccc} \hline\hline $\bfm W_t$ & \multicolumn{4}{c|}{SPCA } & \multicolumn{4}{c|}{SPCA-LS } & \multicolumn{4}{c}{PCA} \\\hline & \multicolumn{4}{c|}{Maturity(Year) } & \multicolumn{4}{c|}{Maturity(Year) } & \multicolumn{4}{c}{Maturity(Year) } \\ & 2 & 3 & 4 & 5 & 2 & 3 & 4 & 5 & 2 & 3 & 4 & 5 \\ $\bfm f_t$ & 44.6 & 43.0 & 38.8 & 37.3 & 41.2 & 39.1 & 35.2 & 34.1 & 30.1 & 25.5 & 23.2 & 21.3 \\ $(\bfm f_t',\ \bfm x_t')'$ & 41.5 & 38.7 & 35.2 & 33.8 & 41.1 & 35.7 & 32.2 & 30.0 & 30.8 & 26.3 & 24.6 & 22.0 \\\hline \end{tabular} \end{center} \end{table}

Simulation Studies

Model settings

We use simulated examples to demonstrate the finite sample performance of the proposed method, which is denoted by SPCA (smoothed PCA), and compare it with SPCA-LS (which uses $\widetilde\bfsym \Sigma$, the least-squares based smoothed PCA, described in Section (ref)) and the benchmark PCA. We set $N=40, \ T=100$ and $K=5$. The supplementary material contains additional simulation results under other $N$ and $T$ combinations, as well as the case of serially dependent data. The findings are similar.

Consider the following data generating process,

eqnarray*[eqnarray* omitted — 243 chars of source]

where $\ensuremath{\boldsymbol{\Lambda}}$ is drawn from i.i.d. standard Normal distribution and $\bfm u_t$ is drawn from either the i.i.d standard Normal distribution or i.i.d. re-scaled Log-Normal distribution $c_1 \{ \exp(1+1.2 \zeta ) - c_2\}$, where $\zeta \sim \mathcal{N}(0, 1)$ and $c_1, c_2 >0$ are chosen such that $u_{it}$ has mean zero and variance $1$.

Here $\tilde{\sigma}(g)$ and $\tilde{\sigma}(\gamma)$ respectively represent the signal and noise levels. Set $\tilde{\sigma}(g)^2+\tilde{\sigma}(\gamma)^2=1$ and $\tilde{\sigma}(g)^2/\tilde{\sigma}(\gamma)^2=\omega$, where $\omega$ controls the ratio between the explained and unexplained parts in the latent factors. To address different signal-noise regimes, we set $\omega=10, 1$ and $0.1$ to represent strong, mild and weak explanatory powers respectively. The baseline $\bfsym \gamma_t^0$ is drawn from i.i.d. standard Normal distribution and the baseline function $\bfm g^0(\cdot)$ is set to be one of the following two models:

enumerate• {\sc Linear model}:\ We set $d=K$ and $\bfm x_t$ is drawn from i.i.d. standard Normal distribution. Let $\bfm g^0(\bfm x_t)= \bfm D \bfm x_t$, where $\bfm D$ is a $K \times K$ matrix with each entry drawn from $U[1,\ 2]$; • {\sc Nonlinear model}:\ We set $d=1$ and $x_t$ is drawn from i.i.d. uniform distribution $[0,1]$. Let $\bfm g^0(x_t)= \{g_1^0(x_t), \cdots , g_K^0(x_t)\}'$ with $g_k^0(x_t)=a_k \cos(2 \pi k x_t) + b_k \sin(2 \pi k x_t)$ for $k=1, \cdots K$. The coefficients $a_k$ and $b_k$ are calibrated from a nonlinear test function $\theta(x)=\sin(x)+2\exp(-30x^2)$ with $x \in [0, 1]$ so that $\bfm g^0$ forms its leading Fourier bases. To save the space, we refer to the example 2 of Dimatteo_01 for the plot of $\theta(x)$.

For each $k\leq K$, we normalize $g_k^0(\bfm x_t)$ and $\gamma^0_{t,k}$ such that they have means zero, and standard deviations one.

Throughout this section, the number of factors is estimated by the eigen-ratio method LamYao,AH. In the following simulated examples, the eigen-ratio method can correctly select $K=5$ in most replications. The sieve basis is chosen as the additive polynomial basis. To account the scale of the noise variance, we also consider the tuning parameter in the Huber loss to admit $\alpha_{T, i}=C_\alpha\tilde{\sigma}_i\sqrt{\frac{T}{\log(NT)}}$, where $\tilde{\sigma}_i=\sqrt{ \frac{1}{T}\sum_t(y_{it}-\tilde E(y_{it}|\bfm x_t))^2}$ and $\tilde E(y_{it}|\bfm x_t)$ is smoothed by sieve least squares using additive polynomial basis of order 5. In Subsection (ref), the tuning parameters $C_{\alpha}$ and $J$ are selected by the in-sample 5-fold cross validation, while in subsection (ref), they are chosen using the out-of-sample 5-fold cross validation.

In-sample Estimation

First, we compare the in-sample model fitting among SPCA, SPCA-LS and PCA under different scenarios. For each scenario, we conduct 200 replications. As the factors and loading may be estimated up to a rotation matrix, the canonical correlations between the parameter and its estimator can be used to measure the estimation accuracy bai03. For Model (I) and (II) we report the sample mean of the median of 5 canonical correlations between the true loading and factors and the estimated ones.

The results are presented in Table (ref). SPCA-LS and SPCA are comparable for light-tail distributions, and are both slightly better than PCA. This implies that we pay little price for the robustness and that the proposed estimators are potentially better than PCA when $N$ is relatively small, due to the merit of the “finite-$N$” asymptotics of the proposed estimators. However, when the error distributions have heavy tails, SPCA yields much better estimation than other methods as expected. SPCA-LS out-performs PCA when $\bfm x_t$ has strong or mild explanatory powers of $\bfm f_t$ which is in line with the discussion in Section (ref). When $\omega = 0.1$, the observed $\bfm x_t$ is not as informative and hence the performance of SPCA and SPCA-LS are close to regular PCA.

Out-of-sample Forecast

We now consider using latent factors in a linear forecast model $ z_{t+1} =\bfsym \beta' \bfm f_t + \epsilon_{t+1}, $ where $\epsilon_{t}$ is drawn from i.i.d. standard normal distribution. For each simulation, the unknown coefficients in $\bfsym \beta$ are independently drawn from uniform distribution $[0.5, \ 1.5]$ to cover a variety of model settings.

We conduct one-step ahead rolling window forecast using the linear model by estimating $\bfsym \beta$ and $\bfm f_t$. The factors are estimated from ((ref)) by SPCA, SPCA-LS or PCA. In each replication, we generate $T+50$ observations in total. For $s=1, \ \cdots \ ,50$, we use the $T$ observations $(z_s,..., z_{T+s-1})$ to forecast $z_{T+s}$. We use PCA as the benchmark and define the relative mean squared error (RMSE) as:

equation*[equation* omitted — 166 chars of source]

where $\widehat{z}_{T+s|T+s-1} $ is the forecast of $z_{T+s}$ based on either SPCA or SPCA-LS while $\tilde{z}_{T+s|T+s-1}^{PCA}$ is the forecast based on PCA.

commentFor SPCA and PCA, they are both based on model (ref) except the factors there are estimated by two different method. For the INT method, the factors are estimated by using SPCA.

For each scenario, we simulate 200 replications and calculate the averaged RMSE as a measurement of the one-step-ahead out-of-sample forecast.

The results are presented in Table (ref). Again, when the tails of error distributions are light, SPCA and SPCA-LS perform comparably. But SPCA outperforms SPCA-LS when the errors have heavy tails. On the other hand, both SPCA and SPCA-LS outperform PCA when $\bfm x_t$ exhibits strong or mild explanatory powers of $\bfm f_t$, but are slightly worse when $\omega$ is small. In general, the SPCA method performs the best under heavy-tailed cases.

table[table omitted — 900 chars of source]

Compare with the interactive effect approach

Here we consider three pairs of sample sizes: $N=40,\ T=150$; $N=60,\ T=100$ and $N=60, \ T=150$. We compare the proposed SPCA method with SPCA-LS (Section (ref)), regular PCA and pure least squares (LS), which models the covariates and estimates the parameters by simply using $$ \min_{\ensuremath{\boldsymbol{\Lambda}},\{\bfm f_t\},\bfsym \beta}\frac{1}{T}\sum_{t=1}^T\|\bfm y_t-\ensuremath{\boldsymbol{\Lambda}} \bfm f_t-\bfm x_t'\bfsym \beta\|^2. $$

In Tables (ref)--(ref), we report sample mean of the median of 5 canonical correlations between the true loading and factors and the estimated ones. Under various sample size combinations, the findings are similar as discussed in Section (ref): (1) both SPCA and SPCA-LS outperform PCA under light-tail distributions when $\bfm x_t$ has strong or mild explanatory powers of $\bfm f_t$; (2) when the error distributions have heavy tails, SPCA outperforms other methods as expected; (3)when $\bfm x_t$ has weak explanatory power, the performance of SPCA and SPCA-LS are close to regular PCA; (4) under all simulated scenarios, the LS approach gives the worst estimation performance.

Serial dependent case

In this subsection, we compare the in-sample model fitting among SPCA, SPCA-LS and PCA under serial dependences. The simulation settings are similar as in Section 5.1 except both $\bfm x_t$ and $\bfsym \gamma_t$ are generated from a stationary VAR(1) model as follows

align*[align* omitted — 162 chars of source]

with $\bfm x_0={\bf 0}$ and $\bfsym \gamma_0= {\bf 0}$. The $(i, j)$th entry of ${\bf \Pi}$ is set to be 0.5 when $i=j$ and $0.4^{|i-j|}$ when $i \neq j$. In addition, $\bfsym \varepsilon_t$ and ${\bf \eta}_t$ are drawn form i.i.d. $N({\bf 0}, \bfm I)$.

The performance under 200 replications are presented in Table (ref) below. Our numerical findings for the independent data continue to hold for serially dependent data: both SPCA and SPCA-LS outperform PCA when $\bfm x_t$ and $\bfm f_t$ are serially correlated. SPCA gives the best performance when the error distributions are heavy-tailed.

table[table omitted — 1,551 chars of source]
comment\begin{table}[htbp] \begin{center} \caption{{\bf In-sample Estimation}: Median of 5 canonical correlations of the estimated loadings/factors and the true ones when $N=40,T=150$: the larger the better } \begin{tabular}{c|c|c|cccc|cccc} \hline\hline & & & \multicolumn{4}{c|}{Model (I)} & \multicolumn{4}{c}{Model (II)} \\ & $\bfm u_t$ & $\omega$ & SPCA & SPCA-LS & PCA & LS & SPCA & SPCA-LS & PCA &LS \\\hline \multirow{6}{*}{Loading}& \multirow{3}{*}{Normal} & 10 & 0.93 & 0.93 & 0.84 & 0.77 & 0.91 & 0.91 & 0.79 & 0.76 \\ & & 1 & 0.90 & 0.90 & 0.84 & 0.78 & 0.86 & 0.87 & 0.79 & 0.76 \\ & & 0.1 & 0.85 & 0.86 & 0.84 & 0.82 & 0.81 & 0.81 & 0.79 & 0.78 \\\cline{2-11} & \multirow{3}{*}{LogN} & 10 & 0.84 & 0.55 & 0.39 & 0.35 & 0.81 & 0.53 & 0.35 & 0.31 \\ & & 1 & 0.80 & 0.51 & 0.39 & 0.37 & 0.78 & 0.49 & 0.35 & 0.32 \\ & & 0.1 & 0.77 & 0.46 & 0.39 & 0.38 & 0.73 & 0.42 & 0.35 & 0.34 \\\hline\hline \multirow{6}{*}{Factors}& \multirow{3}{*}{Normal} & 10 & 0.92 & 0.92 & 0.78 & 0.73 & 0.89 & 0.90 & 0.76 & 0.71 \\ & & 1 & 0.84 & 0.85 & 0.78 & 0.74 & 0.80 & 0.80 & 0.76 & 0.72 \\ & & 0.1 & 0.78 & 0.79 & 0.78 & 0.77 & 0.75 & 0.76 & 0.76 & 0.75 \\\cline{2-11} &\multirow{3}{*}{LogN} & 10 & 0.82 & 0.63 & 0.34 & 0.29 & 0.80 & 0.61 & 0.30 & 0.26 \\ & & 1 & 0.78 & 0.58 & 0.34 & 0.31 & 0.78 & 0.57 & 0.30 & 0.27 \\ & & 0.1 & 0.74 & 0.52 & 0.34 & 0.34 & 0.72 & 0.52 & 0.30 & 0.29 \\\hline \end{tabular} \end{center} \end{table} \begin{table}[htbp] \begin{center} \caption{{\bf In-sample Estimation}: Median of 5 canonical correlations of the estimated loadings/factors and the true ones when $N=60,T=100$: the larger the better } \begin{tabular}{c|c|c|cccc|cccc} \hline\hline & & & \multicolumn{4}{c|}{Model (I)} & \multicolumn{4}{c}{Model (II)} \\ & $\bfm u_t$ & $\omega$ & SPCA & SPCA-LS & PCA & LS & SPCA & SPCA-LS & PCA &LS \\\hline \multirow{6}{*}{Loading}& \multirow{3}{*}{Normal} & 10 & 0.91 & 0.91 & 0.86 & 0.80 & 0.90 & 0.90 & 0.82 & 0.75 \\ & & 1 & 0.89 & 0.89 & 0.86 & 0.81 & 0.86 & 0.86 & 0.82 & 0.77 \\ & & 0.1 & 0.84 & 0.85 & 0.86 & 0.84 & 0.81 & 0.81 & 0.82 & 0.80 \\\cline{2-11} & \multirow{3}{*}{LogN} & 10 & 0.83 & 0.54 & 0.41 & 0.36 & 0.80 & 0.52 & 0.38 & 0.33 \\ & & 1 & 0.79 & 0.50 & 0.41 & 0.39 & 0.76 & 0.47 & 0.38 & 0.34 \\ & & 0.1 & 0.75 & 0.44 & 0.41 & 0.40 & 0.71 & 0.40 & 0.38 & 0.36 \\\hline\hline \multirow{6}{*}{Factors}& \multirow{3}{*}{Normal} & 10 & 0.90 & 0.90 & 0.80 & 0.74 & 0.91 & 0.91 & 0.79 & 0.72 \\ & & 1 & 0.83 & 0.83 & 0.80 & 0.75 & 0.82 & 0.82 & 0.79 & 0.73 \\ & & 0.1 & 0.78 & 0.78 & 0.80 & 0.79 & 0.77 & 0.77 & 0.79 & 0.77 \\\cline{2-11} &\multirow{3}{*}{LogN} & 10 & 0.84 & 0.64 & 0.37 & 0.31 & 0.83 & 0.63 & 0.34 & 0.29 \\ & & 1 & 0.81 & 0.60 & 0.37 & 0.33 & 0.79 & 0.58 & 0.34 & 0.30 \\ & & 0.1 & 0.76 & 0.54 & 0.37 & 0.36 & 0.75 & 0.52 & 0.34 & 0.32 \\\hline \end{tabular} \end{center} \end{table}
table[table omitted — 1,639 chars of source]
table[table omitted — 1,555 chars of source]

Conclusions

We study factor models when the factors depend on observed explanatory characteristics. The proposed method incorporates the explanatory power of these observed covariates, and is robust to possibly heavy-tailed distributions. We focus on the case $\dim(\bfm x_t)$ is finite, and on the rates of convergence for the estimated factors and loadings. Under various signal-noise ratios, substantial improved rates of convergence can be gained.

Related to the above, the idea could be easily extended to the case that $\dim(\bfm x_t)$ is slowly growing (with respect to $(N, T)$). On the other hand, allowing $\dim(\bfm x_t)$ to be fast-growing would require some dimension-reduction treatment combined with covariate selections. In addition, selecting the covariates would be also useful as the quality of the signal is crucial. We shall leave these open questions for future studies.