betaFeedback

WorksMathematics

A Hierarchical Group Factor Model with Characteristic-Driven Global and Local Factors

Abstract

We propose a Hierarchical Group Factor Model (HGFM) for jointly estimating global and group-specific factors with characteristic-driven portfolio weights. Both factor types are constructed from contemporaneous returns, producing endogeneity that we address by quasi-maximum likelihood estimation. Conditional on the interaction matrices, we separate the global and local coefficient components and normalize their factor representations. Under a fixed characteristic design and suitable regularity conditions, we establish local consistency of the coefficient estimators and asymptotic normality for fixed blocks of the normalized factor parameters as both panel dimensions diverge. We also propose an information criterion and establish its consistency for jointly selecting the global rank and the common local rank. In an industry-grouped application to S&P 100 constituents, the HGFM achieves higher explanatory power for realized excess returns and a higher portfolio Sharpe ratio than benchmark asset-pricing models.

Introduction

How can we separate common and group-specific dependence when both global and local factors are constructed from observed responses using characteristic-driven weights?

For stock returns grouped by industry, the distinction is between comovement across industries and additional comovement within each industry. Standard factor models summarize common variation in high-dimensional panels (e.g., [7]), whereas group factor models (GFMs) distinguish global factors shared across groups from local factors specific to each group. Grouped heterogeneity also motivates sparse asset-pricing models [14] and group-specific predictability of mutual fund alphas [15]. Characteristic-based portfolios formed across the full stock panel and within individual industries may capture overlapping variation in returns. This motivates jointly specifying and estimating the global and local components within a common return model. We develop a framework in which predetermined asset characteristics generate time-varying weights for both global and local factors.

Andreou, Gagliardini, Ghysels, and Rubin [3] develop inference for group factor models using within-group principal components and cross-group canonical correlations. We use their common and group-specific latent factor structure as a benchmark. Let ys,ity_{s,it} denote the response variable for the iith cross-sectional unit at time tt in group ss, where s=1,…,Ss = 1,\ldots,S, i=1,…,Nsi = 1,\ldots,N_s, and t=1,…,Tt = 1,\ldots,T. Define Ys,t=(ys,1t,…,ys,Nst)⊤∈RNs\mathbf{Y}_{s,t} = (y_{s,1t},\ldots,y_{s,N_st})^\top\in\mathbb{R}^{N_s} as the vector of response variables for group ss at time tt. The group factor model is then specified as follows:

Ys,t=Bsfg,t+Usfl,s,t+εs,t,(1)\mathbf{Y}_{s,t} = \mathbf{B}_s\mathbf{f}_{g,t} + \mathbf{U}_s\mathbf{f}_{l,s,t} + \boldsymbol{\varepsilon}_{s,t}, \tag*{(1)}

where fg,t∈Rrg\mathbf{f}_{g,t} \in\mathbb{R}^{r_g} and fl,s,t∈Rrl\mathbf{f}_{l,s,t} \in\mathbb{R}^{r_l} are the latent global and local factors, respectively; Bs∈RNs×rg\mathbf{B}_s \in\mathbb{R}^{N_s \times r_g} and Us∈RNs×rl\mathbf{U}_s \in\mathbb{R}^{N_s \times r_l} are their loading matrices; and εs,t∈RNs\boldsymbol{\varepsilon}_{s,t} \in\mathbb{R}^{N_s} is the idiosyncratic error. Thus, fg,t\mathbf{f}_{g,t} affects all groups, whereas fl,s,t\mathbf{f}_{l,s,t} affects only group ss.

The latent-factor specification in (1) leaves the relation between asset characteristics and factor-portfolio weights unspecified. Empirical asset pricing often constructs observable factors by sorting securities on characteristics and forming long–short portfolios, as in the Fama–French factors [17, 18]. The number of such factors has grown large [28], and Feng et al. [23] test whether a new factor adds explanatory power for expected returns beyond the existing set.

We express both types of factors in (1) as linear combinations of contemporaneous returns, with time-varying weights determined by characteristics, and estimate their fixed coefficients jointly. Kelly, Pruitt, and Su [30] use characteristics to parameterize time-varying factor loadings; our characteristics instead determine portfolio weights:

fg,t=Wt−1Yt, fl,s,t=Vs,t−1Ys,t,Wt−1=CgZt−1, Vs,t−1=Cl,sZs,t−1(2)\mathbf{f}_{g,t} = \mathbf{W}_{t-1}\mathbf{Y}_t,\ \mathbf{f}_{l,s,t} = \mathbf{V}_{s,t-1}\mathbf{Y}_{s,t}, \\ \mathbf{W}_{t-1} = \mathbf{C}_g\mathbf{Z}_{t-1},\ \mathbf{V}_{s,t-1} = \mathbf{C}_{l,s}\mathbf{Z}_{s,t-1} \tag*{(2)}

where Yt=(Y1,tT,…,YS,tT)T∈RN\mathbf{Y}_t = (\mathbf{Y}_{1,t}^{\mathsf{T}}, \ldots, \mathbf{Y}_{S,t}^{\mathsf{T}})^{\mathsf{T}} \in\mathbb{R}^{N} and N=∑s=1SNsN = \sum_{s=1}^{S} N_s. The matrices Wt−1∈Rrg×N\mathbf{W}_{t-1} \in\mathbb{R}^{r_g \times N} and Vs,t−1∈Rrl×Ns\mathbf{V}_{s,t-1} \in\mathbb{R}^{r_l \times N_s} are the global and local portfolio weight matrices, respectively, while Cg∈Rrg×K\mathbf{C}_g \in\mathbb{R}^{r_g \times K} and Cl,s∈Rrl×K\mathbf{C}_{l,s} \in\mathbb{R}^{r_l \times K} are fixed characteristic-coefficient matrices. For group ss, Zs,t−1∈RK×Ns\mathbf{Z}_{s,t-1} \in\mathbb{R}^{K \times N_s} contains transformed characteristics available before the period-tt returns are realized, and Zt−1=(Z1,t−1,…,ZS,t−1)∈RK×N\mathbf{Z}_{t-1} = (\mathbf{Z}_{1,t-1}, \ldots, \mathbf{Z}_{S,t-1}) \in\mathbb{R}^{K \times N} collects these characteristics across all groups. The portfolio weights vary over time because Zt−1\mathbf{Z}_{t-1} and Zs,t−1\mathbf{Z}_{s,t-1} vary over time, even though Cg\mathbf{C}_g and Cl,s\mathbf{C}_{l,s} are fixed. This specification restricts each factor’s portfolio weights to the span of the observed characteristics.

By substituting (2) into the group factor model (1), we propose a Hierarchical Group Factor Model (HGFM):

Ys,t=Bs(CgZt−1Yt)+Us(Cl,sZs,t−1Ys,t)+εs,t.(3)\mathbf{Y}_{s,t} = \mathbf{B}_s\left(\mathbf{C}_g\mathbf{Z}_{t-1}\mathbf{Y}_t\right) + \mathbf{U}_s\left(\mathbf{C}_{l,s}\mathbf{Z}_{s,t-1}\mathbf{Y}_{s,t}\right) + \boldsymbol{\varepsilon}_{s,t}. \tag*{(3)}

The global factor in (3) is a portfolio of all assets with weights CgZt−1\mathbf{C}_g\mathbf{Z}_{t-1}, while the local factor for group ss uses only assets in that group, with weights Cl,sZs,t−1\mathbf{C}_{l,s}\mathbf{Z}_{s,t-1}. This linear construction can also represent sorting-based factors (e.g., the Fama–French size and value factors) when the rows of the characteristic matrices encode the stock weights of the constituent portfolios and the coefficient matrices specify their fixed long–short combinations. In our specification, the coefficient matrices are estimated jointly with the other HGFM parameters, allowing the data to determine how predetermined characteristics enter the global and local portfolio weights.

We first separate the global and local coefficient matrices conditional on the interaction matrices, normalize their factor representations, and impose a local information condition for statistical identification. Replacing latent factors with functions of observed returns introduces the outcome YtY_t on the right-hand side of model (3), which gives rise to endogeneity and generally renders ordinary least squares (OLS) estimators inconsistent. We therefore propose a quasi-maximum likelihood estimator (QMLE) computed by gradient ascent with an Armijo line search.

The asymptotic analysis allows both the cross-sectional dimension, NN, and the time dimension, TT, to diverge. Although the coefficient matrices are low-rank, the number of nuisance parameters increases with NN, complicating the analysis of their asymptotic properties [22]. We establish local consistency for the coefficient estimators and consistency and fixed-block asymptotic normality for the normalized factor parameters. The true ranks of the global and local factors are unknown, and standard eigenvalue-ratio methods (e.g., [31, 1]) cannot be directly applied to select them in the present endogenous model. Following Ando and Bai [2], we propose an information criterion (IC) to jointly select the global rank and the common local rank, which consistently recovers the true ranks under suitable regularity assumptions.

Our model is related to Pu, Fang, Lan, Yu, and Zhang (2025), who estimate reduced-rank spatio-temporal coefficient matrices by QMLE with an Armijo line search. Our approach differs in three respects. First, their specification uses an estimated weight matrix multiplying current and lagged returns, whereas ours constructs global and local factors from contemporaneous returns using characteristic-driven, time-varying weights. Second, we jointly estimate the loading and characteristic-coefficient matrices for the global and local components, allowing the characteristic coefficients of local factors to differ across groups while sharing the global factors. Third, our hierarchical specification requires distinct identification conditions and a corresponding theoretical framework. Conditional on the interaction matrices, cross-group characteristic variation separates the global and local coefficient matrices, and empirical characteristic second moments provide the basis for normalizing their factor representations. We also develop an information criterion to jointly select the global rank and the common local rank. Our asymptotic analysis treats the full characteristic design as fixed in advance and takes probability over the errors.

Related literature. The paper relates to three strands of research: group factor models, proxy-assisted factor estimation, and characteristic-based portfolio construction.

The first strand concerns group factor models. Ando and Bai [2] estimate group membership and group-specific latent factors in panels with observed predictors; our model takes group membership as given and estimates characteristic-driven factor weights. Hu, Li, and Wang (2025) identify latent global factors by maximizing their average correlations with group factors, while Choi, Lin, and Shin [12] use cross-group canonical correlations to select latent global factors, with an application to U.S. industries.

Zhang et al. [40] estimate global and local factors in matrix time series, while Choi and Kim [13] use global and national factors for large volatility matrix estimation. Related factor models also cover three-dimensional panels ([29]), matrix-valued time series ([38]), and tensor observations (e.g., [11, 10]). Factor structures of this kind also underlie high-dimensional covariance and portfolio estimation ([16]). The HGFM jointly estimates global and local coefficient matrices that map predetermined characteristics into time-varying portfolio weights, linking both factor components explicitly to observed characteristics and returns.

The second strand uses observable proxies to estimate latent factors. Fan et al. [19] incorporate observed covariates into smoothed principal component analysis (PCA) to recover latent factors, while Wan, Li, Lu, and Song [37] combine responses and a large set of factor proxies through penalized reduced-rank regression to estimate latent factors. Fan and Liao [20] estimate latent factors using diversified cross-sectional projections, with individual characteristics among the proposed sources of portfolio weight matrices. Related studies examine common factor spaces ([4]), the stability of observable–latent factor relations ([34]), and observable-factor risk premia in the presence of omitted factors ([25]). The HGFM specifies both global and local factors as characteristic-weighted contemporaneous responses and jointly estimates the characteristic coefficients and loadings by QMLE to account for the resulting endogeneity.

The third strand concerns characteristic-based factor construction. Kelly et al. [30] use characteristics to parameterize time-varying factor loadings, and Feng, He, Polson, and Xu [24] jointly learn nonlinear characteristic-based portfolio weights and factor loadings by minimizing aggregate realized pricing errors. Brandt et al. [9] parameterize portfolio weights by firm characteristics and estimate their coefficients by maximizing average investor utility. The HGFM specifies separate linear characteristic-based weighting rules for global and group-specific local factors. For this hierarchical specification, we provide identification conditions and QMLE inference for the weight coefficients and loadings.

Simulation and empirical highlights. We examine the finite-sample accuracy of parameter estimation and rank selection through Monte Carlo simulations. We consider independent idiosyncratic errors under Gaussian, standardized-tt, and mixture-normal distributions, as well as cross-sectional correlation and heteroskedasticity. Estimation errors generally decline with TT, including under non-Gaussian errors, and rank-selection accuracy generally improves with TT. Additional simulations examine robustness to sparse characteristic-driven weights, unequal local ranks, larger cross-sectional dimensions, and alternative penalty constants.

Empirically, we analyze daily returns on S&P 100 constituents from five broad industry groups over 2016–2025. The analysis yields three main findings. First, the global factor assigns greater weight to larger stocks and stocks with higher book-to-market ratios, while the characteristic coefficients of local factors differ across industries. Second, combining global and local factors improves realized-return fit across all industry groups relative to either subset alone and the benchmark models. The combined model also explains average returns better than the benchmarks for the full stock panel, although its advantage varies across industries. Third, the combined portfolio achieves the highest annualized Sharpe ratio and the lowest volatility and maximum drawdown among the portfolios considered. Its Sharpe ratio remains above those of the equal- and value-weighted sample portfolios after transaction costs. Additional results examine sensitivity to industry grouping.

The article is structured as follows. § 2 introduces the HGFM, its identification conditions, and the QMLE. § 4 develops the asymptotic theory and rank-selection procedure. § 5 evaluates finite-sample performance by Monte Carlo simulation. § 6 presents the empirical application, and § 7 concludes. The supplementary material contains proofs and additional simulation and empirical results.

The Model and Methodology

In this section, we formally introduce the HGFM. We begin by specifying the model structure, in which the global and local factors are constructed from observed returns using characteristic-driven weights. We then discuss the identification of the global and local components and their factor representations, and detail the QMLE procedure.

Hierarchical Group Factor Model (HGFM)

By introducing global and local factors constructed from observed returns with characteristic-driven weights, we consider the following hierarchical group factor model:

\begin{bmatrix} \mathbf{Y}_{1,t}\\ \mathbf{Y}_{2,t}\\ \vdots\\ \mathbf{Y}_{S,t} \end{bmatrix} = \begin{bmatrix} \mathbf{B}_{1}\\ \mathbf{B}_{2}\\ \vdots\\ \mathbf{B}_{S} \end{bmatrix} \mathbf{C}_{g}\mathbf{Z}_{t-1}\mathbf{Y}_{t} + \begin{bmatrix} \mathbf{U}_{1}\mathbf{C}_{l,1}\mathbf{Z}_{1,t-1} & 0 & \cdots& 0\\ 0 & \mathbf{U}_{2}\mathbf{C}_{l,2}\mathbf{Z}_{2,t-1} & \cdots& 0\\ \vdots& \vdots& \ddots& \vdots\\ 0 & 0 & \cdots& \mathbf{U}_{S}\mathbf{C}_{l,S}\mathbf{Z}_{S,t-1} \end{bmatrix} \begin{bmatrix} \mathbf{Y}_{1,t}\\ \mathbf{Y}_{2,t}\\ \vdots\\ \mathbf{Y}_{S,t} \end{bmatrix} + \begin{bmatrix} \boldsymbol{\varepsilon}_{1,t}\\ \boldsymbol{\varepsilon}_{2,t}\\ \vdots\\ \boldsymbol{\varepsilon}_{S,t} \end{bmatrix}. \tag*{(4)}\end{bmatrix}

In matrix form, Eq. (4) can be rewritten as

Yt=BCgZt−1Yt+bdiag⁡(U1Cl,1Z1,t−1,…,USCl,SZS,t−1)Yt+εt=BCgZt−1Yt+UClbdiag⁡(Z1,t−1,…,ZS,t−1)Yt+εt={AgZt−1+Albdiag⁡(Z1,t−1,…,ZS,t−1)}Yt+εt,(5)\begin{aligned} \mathbf{Y}_{t}=\mathbf{B}\mathbf{C}_{g}\mathbf{Z}_{t-1}\mathbf{Y}_{t}+\operatorname{bdiag}\left(\mathbf{U}_{1}\mathbf{C}_{l,1}\mathbf{Z}_{1,t-1},\ldots,\mathbf{U}_{S}\mathbf{C}_{l,S}\mathbf{Z}_{S,t-1}\right)\mathbf{Y}_{t}+\boldsymbol{\varepsilon}_{t} \\ &=\mathbf{B}\mathbf{C}_{g}\mathbf{Z}_{t-1}\mathbf{Y}_{t}+\mathbf{U}\mathbf{C}_{l}\operatorname{bdiag}\left(\mathbf{Z}_{1,t-1},\ldots,\mathbf{Z}_{S,t-1}\right)\mathbf{Y}_{t}+\boldsymbol{\varepsilon}_{t} \\ &=\left\{\mathbf{A}_{g}\mathbf{Z}_{t-1}+\mathbf{A}_{l}\operatorname{bdiag}\left(\mathbf{Z}_{1,t-1},\ldots,\mathbf{Z}_{S,t-1}\right)\right\}\mathbf{Y}_{t}+\boldsymbol{\varepsilon}_{t}, \tag*{(5)} \end{aligned}

where B=(B1T,…,BST)T∈RN×rg\mathbf{B}=\left(\mathbf{B}_{1}^{\mathsf{T}},\ldots,\mathbf{B}_{S}^{\mathsf{T}}\right)^{\mathsf{T}}\in\mathbb{R}^{N\times r_{g}} denotes the global factor loading matrix, and U=bdiag⁡(U1,…,US)∈RN×Srl\mathbf{U}=\operatorname{bdiag}\left(\mathbf{U}_{1},\ldots,\mathbf{U}_{S}\right)\in\mathbb{R}^{N\times Sr_{l}} denotes the local factor loading matrix. Here, bdiag⁡(⋅)\operatorname{bdiag}(\cdot) forms a block-diagonal matrix from its arguments. Moreover, Cl=bdiag⁡(Cl,1,…,Cl,S)∈RSrl×SK\mathbf{C}_{l}=\operatorname{bdiag}\left(\mathbf{C}_{l,1},\ldots,\mathbf{C}_{l,S}\right)\in\mathbb{R}^{Sr_{l}\times SK} is the block-diagonal matrix of local characteristic coefficients, and \boldsymbol{\varepsilon}_{t}=\left(\boldsymbol{\varepsilon}_{1,t}^{\mathsf{T},\ldots,\boldsymbol{\varepsilon}_{S,t}^{\mathsf{T}}\right)^{\mathsf{T}}} is the idiosyncratic error vector. Define Ag=BCg∈RN×K\mathbf{A}_{g}=\mathbf{B}\mathbf{C}_{g}\in\mathbb{R}^{N\times K}, Al,s=UsCl,s∈RNs×K\mathbf{A}_{l,s}=\mathbf{U}_{s}\mathbf{C}_{l,s}\in\mathbb{R}^{N_{s}\times K}, and Al=UCl=bdiag⁡(Al,1,…,Al,S)∈RN×SK\mathbf{A}_{l}=\mathbf{U}\mathbf{C}_{l}=\operatorname{bdiag}\left(\mathbf{A}_{l,1},\ldots,\mathbf{A}_{l,S}\right)\in\mathbb{R}^{N\times SK}. Define the time-varying coefficient matrix Mt=AgZt−1+bdiag⁡{Al,sZs,t−1}s=1S\mathbf{M}_{t}=\mathbf{A}_{g}\mathbf{Z}_{t-1}+\operatorname{bdiag}\left\{\mathbf{A}_{l,s}\mathbf{Z}_{s,t-1}\right\}_{s=1}^{S}. Then model (5) is Yt=MtYt+εt\mathbf{Y}_{t}=\mathbf{M}_{t}\mathbf{Y}_{t}+\boldsymbol{\varepsilon}_{t}. We refer to Mt\mathbf{M}_{t} as the interaction matrix because it describes contemporaneous dependence among the observed units. When KK, rgr_{g}, rlr_{l}, and SS are fixed, the number of unknown parameters associated with Ag\mathbf{A}_{g} and Al\mathbf{A}_{l} grows linearly with NN. This linear growth follows from the fixed dimension of the characteristics, while the low-rank restrictions further constrain the global and local coefficient matrices. Under suitable conditions, the model can be consistently estimated as TT diverges.

Remark 1 (Relation to spatial models). Our model (5) is related to spatial autoregressive (SAR) models, as both capture contemporaneous dependence. Traditional SAR models rely on a prespecified spatial weight matrix, whereas the HGFM uses predetermined characteristics and estimated low-rank coefficient matrices to construct time-varying interaction matrices. The model decomposes these interactions into global and local components. Setting B=0\mathbf{B}=0 leaves only within-group interactions, while setting U=0\mathbf{U}=0 yields a purely global structure. The model therefore represents dependence both across and within groups through observed characteristics.

Model Identifiability

Identification issues commonly arise in latent factor models [5, 6, 8, 39]. The proposed model raises two identification issues: the separation of the global and group-specific components and the factorization of each low-rank coefficient matrix. We address these two issues sequentially. First, we exploit cross-group characteristic variation to separate the global and local coefficient matrices. We then impose a Singular Value Decomposition (SVD) normalization on the low-rank factors of each separated coefficient matrix.

The first identification issue concerns whether the time-invariant global and local coefficient matrices can be separately recovered from the time-varying interaction matrix. To illustrate this issue, let xs,t=Zs,t−1Ys,t\mathbf{x}_{s,t}=\mathbf{Z}_{s,t-1}\mathbf{Y}_{s,t} and x−s,t=∑j≠sZj,t−1Yj,t\mathbf{x}_{-s,t}=\sum_{j\ne s}\mathbf{Z}_{j,t-1}\mathbf{Y}_{j,t}, and let Ag,s\mathbf{A}_{g,s} denote the row block of Ag\mathbf{A}_g corresponding to group ss. The model for group ss can then be written as Ys,t=Ag,sx−s,t+(Ag,s+Al,s)xs,t+εs,t\mathbf{Y}_{s,t}=\mathbf{A}_{g,s}\mathbf{x}_{-s,t}+(\mathbf{A}_{g,s}+\mathbf{A}_{l,s})\mathbf{x}_{s,t}+\boldsymbol{\varepsilon}_{s,t}. Thus, the cross-group input x−s,t\mathbf{x}_{-s,t} is governed exclusively by Ag,s\mathbf{A}_{g,s}, whereas the own-group input xs,t\mathbf{x}_{s,t} is governed by Ag,s+Al,s\mathbf{A}_{g,s}+\mathbf{A}_{l,s}. Given identification of the time-varying interaction matrix, this structure allows us to distinguish Ag,s\mathbf{A}_{g,s} and Al,s\mathbf{A}_{l,s} when the characteristics vary sufficiently both within and outside group ss.

Unless otherwise stated, we treat the entire characteristic array as a nonstochastic design fixed in advance throughout the theoretical analysis. For each group ss, let Z−s,t−1∈RK×(N−Ns)\mathbf{Z}_{-s,t-1} \in\mathbb{R}^{K \times(N-N_s)} denote the concatenation of {Zj,t−1:j≠s}\{\mathbf{Z}_{j,t-1}: j \ne s\}, and define the corresponding design matrices by

QZ,s,T=1TNs∑t=1TZs,t−1Zs,t−1T,QZ,−s,T=1T(N−Ns)∑t=1TZ−s,t−1Z−s,t−1T.\mathbf{Q}_{Z,s,T} = \frac{1}{TN_s} \sum_{t=1}^{T} \mathbf{Z}_{s,t-1}\mathbf{Z}_{s,t-1}^{\mathsf{T}}, \qquad\mathbf{Q}_{Z,-s,T} = \frac{1}{T(N-N_s)} \sum_{t=1}^{T} \mathbf{Z}_{-s,t-1}\mathbf{Z}_{-s,t-1}^{\mathsf{T}}.

The following condition ensures that the characteristic variation is nondegenerate both within group ss and across the groups outside ss.

Condition 1. There exist constants 0<cZ<CZ<∞0 < c_Z < C_Z < \infty such that, for every s=1,…,Ss = 1,\ldots,S,

cZ≤λmin⁡(QZ,s,T)≤λmax⁡(QZ,s,T)≤CZ,c_Z \le\lambda_{\min}(\mathbf{Q}_{Z,s,T}) \le\lambda_{\max}(\mathbf{Q}_{Z,s,T}) \le C_Z,
cZ≤λmin⁡(QZ,−s,T)≤λmax⁡(QZ,−s,T)≤CZ.c_Z \le\lambda_{\min}(\mathbf{Q}_{Z,-s,T}) \le\lambda_{\max}(\mathbf{Q}_{Z,-s,T}) \le C_Z.

Condition 1 rules out degenerate characteristic directions that would prevent the global and local components from being distinguished. The following proposition states that if two collections generate the same interaction matrices at the design dates, their global and group-specific coefficient matrices coincide.

Proposition 1. Suppose that Condition 1 holds. Consider two collections of coefficient matrices {Ag,Al,s:s=1,…,S}\{\mathbf{A}_g, \mathbf{A}_{l,s}: s = 1,\ldots,S\} and {A~g,A~l,s:s=1,…,S}\{\widetilde{\mathbf{A}}_g, \widetilde{\mathbf{A}}_{l,s}: s = 1,\ldots,S\}. For every t=1,…,Tt = 1,\ldots,T, if

AgZt−1+bdiag⁡{Al,sZs,t−1}s=1S=A~gZt−1+bdiag⁡{A~l,sZs,t−1}s=1S,\mathbf{A}_g\mathbf{Z}_{t-1} + \operatorname{bdiag}\{\mathbf{A}_{l,s}\mathbf{Z}_{s,t-1}\}_{s=1}^{S} = \widetilde{\mathbf{A}}_g\mathbf{Z}_{t-1} + \operatorname{bdiag}\{\widetilde{\mathbf{A}}_{l,s}\mathbf{Z}_{s,t-1}\}_{s=1}^{S},

then Ag=A~g\mathbf{A}_g = \widetilde{\mathbf{A}}_g and Al,s=A~l,s\mathbf{A}_{l,s} = \widetilde{\mathbf{A}}_{l,s} for every s=1,…,Ss = 1,\ldots,S.

Under Condition 1 and conditional on the interaction matrices, Proposition 1 separates Ag,s\mathbf{A}_{g,s} through cross-group characteristic variation and then Al,s\mathbf{A}_{l,s} through within-group variation, for every s=1,…,Ss = 1,\ldots,S.

The second identification issue is the usual rotation indeterminacy of low-rank factor models [6, 35]. For any nonsingular matrices Hg\mathbf{H}_g and Hl,sH_{l,s}, the transformations B~=BHg\widetilde{B}=BH_g, C~g=Hg−1Cg\widetilde{C}_g=H_g^{-1}C_g, U~s=UsHl,s\widetilde{U}_s=U_sH_{l,s}, and C~l,s=Hl,s−1Cl,s\widetilde{C}_{l,s}=H_{l,s}^{-1}C_{l,s} leave BCgBC_g and UsCl,sU_sC_{l,s} unchanged. Thus, the factor representations remain unidentified even after the global and local coefficient matrices have been separated. We therefore impose an SVD normalization to identify the factors without changing the corresponding coefficient matrices.

Define the empirical second-moment matrices of the observed characteristics as

ΓZ,T=1T∑t=1TZt−1Zt−1T,ΓZ,s,T=1T∑t=1TZs,t−1Zs,t−1T.\Gamma_{Z,T}=\frac{1}{T}\sum_{t=1}^{T}Z_{t-1}Z_{t-1}^{\mathsf{T}}, \qquad\Gamma_{Z,s,T}=\frac{1}{T}\sum_{t=1}^{T}Z_{s,t-1}Z_{s,t-1}^{\mathsf{T}}.

Since ΓZ,T=∑s=1SNsQZ,s,T\Gamma_{Z,T}=\sum_{s=1}^{S}N_sQ_{Z,s,T} and ΓZ,s,T=NsQZ,s,T\Gamma_{Z,s,T}=N_sQ_{Z,s,T}, both matrices are positive definite under Condition 1. Using these matrices, the following normalization condition identifies each low-rank factorization up to paired sign changes.

Condition 2. The global and local factor matrices satisfy BTB=CgΓZ,TCgT=DgB^{\mathsf{T}}B=C_g\Gamma_{Z,T}C_g^{\mathsf{T}}=D_g and UsTUs=Cl,sΓZ,s,TCl,sT=Dl,sU_s^{\mathsf{T}}U_s=C_{l,s}\Gamma_{Z,s,T}C_{l,s}^{\mathsf{T}}=D_{l,s} for every s=1,…,Ss=1,\ldots,S, where DgD_g and Dl,sD_{l,s} are diagonal matrices whose diagonal entries are positive, distinct, and arranged in decreasing order.

Condition 2 adopts the standard balanced normalization used in factor models ([6, 8, 39, 35]). The balancing equality fixes relative scale, while diagonality, distinctness, and decreasing ordering remove rotation and permutation indeterminacies, leaving only paired sign changes.

Proposition 2. Suppose that Ag=BCg=B~C~gA_g=BC_g=\widetilde{B}\widetilde{C}_g and Al,s=UsCl,s=U~sC~l,sA_{l,s}=U_sC_{l,s}=\widetilde{U}_s\widetilde{C}_{l,s} for every s=1,…,Ss=1,\ldots,S. If both representations satisfy Condition 2, then there exist diagonal sign matrices SgS_g and Sl,sS_{l,s}, with diagonal entries in {−1,1}\{-1,1\}, such that

B~=BSg,C~g=SgCg,U~s=UsSl,s,C~l,s=Sl,sCl,s.\widetilde{B}=BS_g,\qquad\widetilde{C}_g=S_gC_g,\qquad\widetilde{U}_s=U_sS_{l,s},\qquad\widetilde{C}_{l,s}=S_{l,s}C_{l,s}.

Proposition 2 establishes the uniqueness of the normalized factorization up to paired sign changes. To complete the factorization argument for coefficient matrices with the specified exact ranks and singular-value gaps, we construct a normalized representation by an ordered singular value decomposition.

Proposition 3. Suppose that Condition 1 holds. Assume that rank⁡(Ag)=rg\operatorname{rank}(\mathbf{A}_g)=r_g and rank⁡(Al,s)=rl\operatorname{rank}(\mathbf{A}_{l,s})=r_l for every ss. Assume that the positive singular values of AgΓZ,T1/2\mathbf{A}_g\mathbf{\Gamma}_{Z,T}^{1/2} are distinct and that, for every s=1,…,Ss=1,\ldots,S, those of Al,sΓZ,s,T1/2\mathbf{A}_{l,s}\mathbf{\Gamma}_{Z,s,T}^{1/2} are also distinct. Let AgΓZ,T1/2=PgΛgVg⊤\mathbf{A}_g\mathbf{\Gamma}_{Z,T}^{1/2}=\mathbf{P}_g\mathbf{\Lambda}_g\mathbf{V}_g^{\top} be its ordered thin singular value decomposition, and define B†=PgΛg1/2\mathbf{B}^{\dagger}=\mathbf{P}_g\mathbf{\Lambda}_g^{1/2} and Cg†=Λg1/2Vg⊤ΓZ,T−1/2\mathbf{C}_g^{\dagger}=\mathbf{\Lambda}_g^{1/2}\mathbf{V}_g^{\top}\mathbf{\Gamma}_{Z,T}^{-1/2}. For each ss, let Al,sΓZ,s,T1/2=Pl,sΛl,sVl,s⊤\mathbf{A}_{l,s}\mathbf{\Gamma}_{Z,s,T}^{1/2}=\mathbf{P}_{l,s}\mathbf{\Lambda}_{l,s}\mathbf{V}_{l,s}^{\top} be its ordered thin singular value decomposition, and define Us†=Pl,sΛl,s1/2\mathbf{U}_s^{\dagger}=\mathbf{P}_{l,s}\mathbf{\Lambda}_{l,s}^{1/2} and Cl,s†=Λl,s1/2Vl,s⊤ΓZ,s,T−1/2\mathbf{C}_{l,s}^{\dagger}=\mathbf{\Lambda}_{l,s}^{1/2}\mathbf{V}_{l,s}^{\top}\mathbf{\Gamma}_{Z,s,T}^{-1/2}. Then B†Cg†=Ag\mathbf{B}^{\dagger}\mathbf{C}_g^{\dagger}=\mathbf{A}_g and Us†Cl,s†=Al,s\mathbf{U}_s^{\dagger}\mathbf{C}_{l,s}^{\dagger}=\mathbf{A}_{l,s} for every s=1,…,Ss=1,\ldots,S. Moreover, the constructed factors satisfy Condition 2 with Dg=Λg\mathbf{D}_g=\mathbf{\Lambda}_g and Dl,s=Λl,s\mathbf{D}_{l,s}=\mathbf{\Lambda}_{l,s}.

For coefficient matrices of the stated exact ranks and with distinct positive singular values, Proposition 3 constructs the normalized factorization and Proposition 2 gives uniqueness up to paired signs. Together with Proposition 1, these results establish the uniqueness of the coefficient decomposition and its normalized factorization for the given interaction matrices.

Remark 2 (Equivariance to the ordering of the observed units). If assets are reordered within the known groups, the responses and columns of the characteristic matrices must be reordered jointly. If the groups are reordered, the group partition must be permuted as well. The rows of the loading matrices follow the same reordering as the assets. The characteristic coefficients and factor series are preserved when the same sign convention is used. This is because jointly reordering returns and characteristics leaves their product unchanged. The normalization therefore does not depend on which assets appear first and does not require anchor assets.

Quasi-maximum Likelihood Estimator

Because Yt\mathbf{Y}_t appears on both sides of model (5), the contemporaneous regressors depend on the errors, so ordinary least squares is generally inconsistent. The ranks rgr_g and rlr_l are treated as known throughout this subsection. We therefore adopt the quasi-maximum likelihood approach of Lee [32]. Let θ=(B,Cg,{Us,Cl,s}s=1S\boldsymbol{\theta}=(\mathbf{B},\mathbf{C}_g,\{\mathbf{U}_s,\mathbf{C}_{l,s}\}_{s=1}^{S} collect the factor parameters before normalization, and let τ=(τ1,…,τN)⊤\boldsymbol{\tau}=(\tau_1,\ldots,\tau_N)^\top, τi=log⁡σi2\tau_i=\log\sigma_i^2, and Φ(τ)=diag⁡(eτ1,…,eτN)\boldsymbol{\Phi}(\boldsymbol{\tau})=\operatorname{diag}(e^{\tau_1},\ldots,e^{\tau_N}). The associated coefficient parameter is η(θ,τ)=(BCg,{UsCl,s}s=1S,τ)\boldsymbol{\eta}(\boldsymbol{\theta},\boldsymbol{\tau})=(\mathbf{B}\mathbf{C}_g,\{\mathbf{U}_s\mathbf{C}_{l,s}\}_{s=1}^S,\boldsymbol{\tau}). Let Mt(θ)\mathbf{M}_t(\boldsymbol{\theta}) be the interaction matrix in model (5) generated by BCg\mathbf{B}\mathbf{C}_g and {UsCl,s}s=1S\{\mathbf{U}_s\mathbf{C}_{l,s}\}_{s=1}^S, and define St(θ)=IN−Mt(θ)\mathbf{S}_t(\boldsymbol{\theta})=\mathbf{I}_N-\mathbf{M}_t(\boldsymbol{\theta}). Up to an additive constant independent of the parameters, the sample-size-normalized negative quasi-log-likelihood is

QT,F(θ,τ)=12N∑i=1Nτi−1NT∑t=1Tlog⁡∣St(θ)∣+12NT∑t=1TYt⊤St(θ)⊤Φ(τ)−1St(θ)Yt.(6)Q_{T,F}(\boldsymbol{\theta},\boldsymbol{\tau})=\frac{1}{2N}\sum_{i=1}^{N}\tau_i-\frac{1}{NT}\sum_{t=1}^{T}\log\lvert\mathbf{S}_t(\boldsymbol{\theta})\rvert+\frac{1}{2NT}\sum_{t=1}^{T}\mathbf{Y}_t^\top\mathbf{S}_t(\boldsymbol{\theta})^\top\boldsymbol{\Phi}(\boldsymbol{\tau})^{-1}\mathbf{S}_t(\boldsymbol{\theta})\mathbf{Y}_t. \tag*{(6)}

Let VN(rg,rl)\mathcal{V}_N(r_g,r_l) denote the set of (θ,τ)(\boldsymbol{\theta},\boldsymbol{\tau}) whose coefficient products and variance vector satisfy the compactness, rank, and stability restrictions in Condition 3. Estimation proceeds by minimizing QT,FQ_{T,F} over this parameter space.

The variance parameters can be concentrated out of the criterion. For fixed θ\boldsymbol{\theta}, the variance first-order conditions give

σ^i2(θ)=1T∑t=1T{(ιi(N))⊤St(θ)Yt}2,i=1,…,N,(7)\widehat{\sigma}_i^2(\boldsymbol{\theta})=\frac{1}{T}\sum_{t=1}^{T}\left\{(\boldsymbol{\iota}_i^{(N)})^\top\mathbf{S}_t(\boldsymbol{\theta})\mathbf{Y}_t\right\}^2,\qquad i=1,\ldots,N, \tag*{(7)}

where ιj(d)\boldsymbol{\iota}_j^{(d)} denotes the jjth canonical vector in Rd\mathbb{R}^d. Substituting (7) into (6), multiplying by −NT-NT, and omitting the resulting constant gives the concentrated quasi-log-likelihood

LT,c(θ)=∑t=1Tlog⁡∣St(θ)∣−T2∑i=1Nlog⁡σ^i2(θ).(8)\mathcal{L}_{T,c}(\boldsymbol{\theta})=\sum_{t=1}^{T}\log\lvert\mathbf{S}_t(\boldsymbol{\theta})\rvert-\frac{T}{2}\sum_{i=1}^{N}\log\widehat{\sigma}_i^2(\boldsymbol{\theta}). \tag*{(8)}

Following He et al. (2023) and Pu et al. (2025), we maximize LT,c\mathcal{L}_{T,c} over the factor matrices using a gradient-based algorithm with an Armijo line search. Denote the retained, not-yet-normalized factor matrices by θ~\widetilde{\boldsymbol{\theta}}. Neither singular-value truncation nor the SVD normalization is imposed during the iterations. In implementation, the optimization is repeated from multiple full-rank initial values, and the candidate attaining the largest value of LT,c\mathcal{L}_{T,c} is retained. The corresponding log-variance estimates are τ^i=log⁡σ^i2(θ~)\widehat{\tau}_i=\log\widehat{\sigma}_i^2(\widetilde{\boldsymbol{\theta}}).

After optimization, define A^g=B~C~g\widehat{\mathbf{A}}_g=\widetilde{\mathbf{B}}\widetilde{\mathbf{C}}_g and A^l,s=U~sC~l,s\widehat{\mathbf{A}}_{l,s}=\widetilde{\mathbf{U}}_s\widetilde{\mathbf{C}}_{l,s}. We then apply the post-estimation normalization in Proposition 3 to these estimated coefficient matrices, using ΓZ,T\Gamma_{Z,T} and ΓZ,s,T\Gamma_{Z,s,T} for the global and local components, respectively, obtaining B^\widehat{B}, C^g\widehat{C}_g, U^s\widehat{U}_s, and C^l,s\widehat{C}_{l,s}. By construction, B^C^g=A^g\widehat{B}\widehat{C}_g=\widehat{A}_g and U^sC^l,s=A^l,s\widehat{U}_s\widehat{C}_{l,s}=\widehat{A}_{l,s} for every s=1,…,Ss=1,\ldots,S. Consequently, the normalization leaves the estimated coefficient matrices, variance estimates, and quasi-likelihood value unchanged while selecting an identified factor representation.

Algorithm 1: Gradient algorithm with Armijo line search

Input: Group returns {Ys,t}\{Y_{s,t}\}; predetermined characteristics {Zs,t−1}\{Z_{s,t-1}\}; ranks rg,rlr_g,r_l; MM initial values {θ(0,m)}m=1M\{\theta^{(0,m)}\}_{m=1}^{M}, where θ=(B,Cg,{Us,Cl,s}s=1S)\theta=(B,C_g,\{U_s,C_{l,s}\}_{s=1}^{S}); maximum numbers of iterations Kout,KinK_{\mathrm{out}},K_{\mathrm{in}}; scale factor ϖ\varpi; tolerance χ\chi; Armijo parameter ξ\xi.

Output: σ^2,B^,C^g,{U^s,C^l,s}s=1S\widehat{\sigma}^{2},\widehat{B},\widehat{C}_g,\{\widehat{U}_s,\widehat{C}_{l,s}\}_{s=1}^{S}.

  1. for m=1,…,Mm=1,\ldots,M do

  2. for k=0,…,Kout−1k=0,\ldots,K_{\mathrm{out}}-1 do

  3. Compute σ^2(k,m)\widehat{\sigma}^{2(k,m)} from (7), LT,c(θ(k,m))\mathcal{L}_{T,c}(\theta^{(k,m)}), and g(k,m)=∇θLT,c(θ(k,m))g^{(k,m)}=\nabla_{\theta}\mathcal{L}_{T,c}(\theta^{(k,m)});

  4. if ∥g(k,m)∥∞<χ\lVert g^{(k,m)}\rVert_{\infty}<\chi then

  5. Mark the current estimate as converged and break;

  6. end

  7. for i=1,…,Kini=1,\ldots,K_{\mathrm{in}} do

  8. Set ai=ϖi−1a_i=\varpi^{i-1} and θi(k+1,m)=θ(k,m)+aig(k,m)\theta_i^{(k+1,m)}=\theta^{(k,m)}+a_i g^{(k,m)};

  9. if LT,c(θi(k+1,m))≥LT,c(θ(k,m))+ξai∥g(k,m)∥22\mathcal{L}_{T,c}(\theta_i^{(k+1,m)})\geq\mathcal{L}_{T,c}(\theta^{(k,m)})+\xi a_i\lVert g^{(k,m)}\rVert_2^2 then

  10. Set θ(k+1,m)=θi(k+1,m)\theta^{(k+1,m)}=\theta_i^{(k+1,m)} and break;

  11. end

  12. end

  13. end

  14. end

  15. Among the admissible converged terminal estimates, retain the one with the largest profile likelihood; denote its factors by θ~\widetilde{\theta};

  16. Apply the normalization in Proposition 3 to the retained coefficient matrices B~C~g\widetilde{B}\widetilde{C}_g and U~sC~l,s\widetilde{U}_s\widetilde{C}_{l,s}, using ΓZ,T\Gamma_{Z,T} and ΓZ,s,T\Gamma_{Z,s,T}, respectively, and denote the normalized factors by B^,C^g,{U^s,C^l,s}s=1S\widehat{B},\widehat{C}_g,\{\widehat{U}_s,\widehat{C}_{l,s}\}_{s=1}^{S};

  17. return σ^2,B^,C^g,{U^s,C^l,s}s=1S\widehat{\sigma}^{2},\widehat{B},\widehat{C}_g,\{\widehat{U}_s,\widehat{C}_{l,s}\}_{s=1}^{S};

Remark 3 (Scope of the numerical guarantee). Algorithm 1 produces numerical candidate estimates from MM initial values. The Armijo line search increases the objective along accepted updates, but neither the line search nor multiple starts guarantees a global optimum. The theory below establishes the existence and rate of a local solution in a neighborhood of the identified parameter.

Theoretical Results

Theoretical Results

Under the stated conditions, we establish local consistency for the coefficient and log-variance estimators, transfer this rate to the normalized factor representation, and derive asymptotic normality for prespecified loading rows and characteristic-coefficient columns. We then establish consistency of joint global and common local rank selection under suitable penalty rates.

Average Consistency

Before establishing the average consistency of the proposed QMLE, we introduce some notation. For a generic matrix A=(aij)∈Rm×n\mathbf{A} = (a_{ij}) \in\mathbb{R}^{m \times n}, vec⁡(A)\operatorname{vec}(\mathbf{A}) denotes its column-wise vectorization. Let λmin⁡(A)\lambda_{\min}(\mathbf{A}) and λmax⁡(A)\lambda_{\max}(\mathbf{A}) denote the minimum and maximum eigenvalues of a symmetric matrix A\mathbf{A}, respectively. In addition, ∥A∥q\lVert\mathbf{A}\rVert_q denotes the induced matrix qq-norm for q=1,2,∞q = 1,2,\infty, and ∥A∥F=(∑i=1m∑j=1naij2)1/2\lVert\mathbf{A}\rVert_F = \left(\sum_{i=1}^{m}\sum_{j=1}^{n}a_{ij}^{2}\right)^{1/2} denotes the Frobenius norm.

Let η=(Ag,{Al,s}s=1S,τ)\boldsymbol{\eta} = (\mathbf{A}_g,\{\mathbf{A}_{l,s}\}_{s=1}^{S},\boldsymbol{\tau}) denote the coefficient and log-variance parameters, and let ηv\boldsymbol{\eta}_v denote the column-wise vectorization of (Ag,Al,1,…,Al,S,τ)(\mathbf{A}_g,\mathbf{A}_{l,1},\ldots,\mathbf{A}_{l,S},\boldsymbol{\tau}). The corresponding scaling matrix is Dη,N=bdiag⁡{INK,{INsK}s=1S,NIN}\mathbf{D}_{\eta,N} = \operatorname{bdiag}\{\mathbf{I}_{NK},\{\mathbf{I}_{N_sK}\}_{s=1}^{S},N\mathbf{I}_N\}. For a vector v\mathbf{v} with the same dimension as ηv\boldsymbol{\eta}_v, set ∥v∥N2=vTDη,N−1v\lVert\mathbf{v}\rVert_N^2 = \mathbf{v}^{\mathsf{T}}\mathbf{D}_{\eta,N}^{-1}\mathbf{v}. For two coefficient-space parameters ηa\boldsymbol{\eta}_a and ηb\boldsymbol{\eta}_b, use the same notation for their vectorized difference:

∥ηa−ηb∥N2=∥Ag,a−Ag,b∥F2+∑s=1S∥Al,s,a−Al,s,b∥F2+N−1∥τa−τb∥22.\lVert\boldsymbol{\eta}_a-\boldsymbol{\eta}_b\rVert_N^2 = \lVert\mathbf{A}_{g,a}-\mathbf{A}_{g,b}\rVert_F^2+\sum_{s=1}^{S}\lVert\mathbf{A}_{l,s,a}-\mathbf{A}_{l,s,b}\rVert_F^2+N^{-1}\lVert\boldsymbol{\tau}_a-\boldsymbol{\tau}_b\rVert_2^2.

Equivalently, ∥ηa−ηb∥N2=(ηv,a−ηv,b)TDη,N−1(ηv,a−ηv,b)\lVert\boldsymbol{\eta}_a-\boldsymbol{\eta}_b\rVert_N^2=(\boldsymbol{\eta}_{v,a}-\boldsymbol{\eta}_{v,b})^{\mathsf{T}}\mathbf{D}_{\eta,N}^{-1}(\boldsymbol{\eta}_{v,a}-\boldsymbol{\eta}_{v,b}).

For average consistency, we use the coefficient parameter η\boldsymbol{\eta}. Define St(η)=IN−Mt(η)\mathbf{S}_t(\boldsymbol{\eta})=\mathbf{I}_N-\mathbf{M}_t(\boldsymbol{\eta}), where Mt(η)\mathbf{M}_t(\boldsymbol{\eta}) is the interaction matrix in (5) written in terms of Ag\mathbf{A}_g and {Al,s}s=1S\{\mathbf{A}_{l,s}\}_{s=1}^{S}.

Since QT,FQ_{T,F} depends on the direct factors only through Ag=BCg\mathbf{A}_g=\mathbf{B}\mathbf{C}_g and Al,s=UsCl,s\mathbf{A}_{l,s}=\mathbf{U}_s\mathbf{C}_{l,s}, its equivalent coefficient-space form is

QT(η)=12N∑i=1Nτi−1NT∑t=1Tlog⁡∣St(η)∣+12NT∑t=1TYtTSt(η)TΦ(τ)−1St(η)Yt.(9)Q_T(\eta) = \frac{1}{2N}\sum_{i=1}^{N}\tau_i - \frac{1}{NT}\sum_{t=1}^{T}\log\lvert S_t(\eta)\rvert + \frac{1}{2NT}\sum_{t=1}^{T}\mathbf{Y}_t^{\mathsf{T}}S_t(\eta)^{\mathsf{T}}\Phi(\tau)^{-1}S_t(\eta)\mathbf{Y}_t. \tag*{(9)}

The admissible coefficient space HN(rg,rl)\mathcal{H}_N(r_g,r_l) is the compact, rank-restricted, and stable set specified in Condition 3 below. Define the population criterion by Q(η)=E{QT(η)}Q(\eta) = \mathbb{E}\{Q_T(\eta)\} on the admissible coefficient space. Let η0\eta^0 denote the data-generating parameter; it minimizes the population criterion.

We next state the regularity conditions.

Condition 3. Let HN(rg,rl)\mathcal{H}_N(r_g,r_l) be a compact parameter space for η=(Ag,{Al,s}s=1S,τ)\eta= (\mathbf{A}_g,\{\mathbf{A}_{l,s}\}_{s=1}^{S},\tau), with rank⁡(Ag)≤rg\operatorname{rank}(\mathbf{A}_g) \le r_g and rank⁡(Al,s)≤rl\operatorname{rank}(\mathbf{A}_{l,s}) \le r_l for each ss. There exist constants 0<CA,CS<∞0 < C_A,C_S < \infty, such that, for all η∈HN(rg,rl)\eta\in\mathcal{H}_N(r_g,r_l),

N∥Ag∥F2+∑s=1SNs∥Al,s∥F2≤CA,∥τ∥∞≤CA.N\lVert\mathbf{A}_g\rVert_F^2 + \sum_{s=1}^{S}N_s\lVert\mathbf{A}_{l,s}\rVert_F^2 \le C_A,\qquad\lVert\tau\rVert_\infty\le C_A.

Moreover, St(η)\mathbf{S}_t(\eta) is nonsingular for every t=1,…,Tt = 1,\ldots,T, and

sup⁡η∈HN(rg,rl)max⁡1≤t≤T{∥St(η)∥1+∥St(η)∥∞+∥St(η)−1∥1+∥St(η)−1∥∞}≤CS.\sup_{\eta\in\mathcal{H}_N(r_g,r_l)}\max_{1\le t\le T}\left\{\lVert\mathbf{S}_t(\eta)\rVert_1+\lVert\mathbf{S}_t(\eta)\rVert_\infty+\lVert\mathbf{S}_t(\eta)^{-1}\rVert_1+\lVert\mathbf{S}_t(\eta)^{-1}\rVert_\infty\right\}\le C_S.

Condition 4. The true coefficient matrices have ranks rgr_g and rlr_l, and η0\eta^0 is uniformly interior to the non-rank restrictions in Condition 3. At the normalized factor representation, let Dg0=(B0)TB0=Cg0ΓZ,T(Cg0)T\mathbf{D}_g^0=(\mathbf{B}^0)^{\mathsf{T}}\mathbf{B}^0=\mathbf{C}_g^0\Gamma_{Z,T}(\mathbf{C}_g^0)^{\mathsf{T}} and write Dg0=diag⁡(dg,1,…,dg,rg)\mathbf{D}_g^0=\operatorname{diag}(d_{g,1},\ldots,d_{g,r_g}). For each ss, let Dl,s0=(Us0)TUs0=Cl,s0ΓZ,s,T(Cl,s0)T\mathbf{D}_{l,s}^0=(\mathbf{U}_s^0)^{\mathsf{T}}\mathbf{U}_s^0=\mathbf{C}_{l,s}^0\Gamma_{Z,s,T}(\mathbf{C}_{l,s}^0)^{\mathsf{T}} and write Dl,s0=diag⁡(dl,s,1,…,dl,s,rl)\mathbf{D}_{l,s}^0=\operatorname{diag}(d_{l,s,1},\ldots,d_{l,s,r_l}). There exist constants 0<cD<CD<∞0<c_D<C_D<\infty and cgap>0c_{\mathrm{gap}}>0, independent of N,TN,T, such that every diagonal entry of these matrices lies in [cD,CD][c_D,C_D] and every adjacent difference within each matrix is at least cgapc_{\mathrm{gap}}.

Condition 5. The number of groups S≥2S\ge2 is fixed, and there exist constants 0<cgrp<Cgrp<∞0<c_{\mathrm{grp}}<C_{\mathrm{grp}}<\infty such that cgrp≤Ns/N≤Cgrpc_{\mathrm{grp}}\le N_s/N\le C_{\mathrm{grp}} for every s=1,…,Ss=1,\ldots,S.

Condition 6. For some constant Cchar<∞C_{\mathrm{char}} < \infty, the characteristic array satisfies

sup⁡N,Tmax⁡1≤t≤Tmax⁡1≤k≤K, 1≤i≤N∣Zki,t−1∣≤Cchar.\sup_{N,T}\max_{1\le t\le T}\max_{1\le k\le K,\,1\le i\le N}|Z_{ki,t-1}| \le C_{\mathrm{char}}.

The error vectors are independent over time, and their entries are mutually independent within each date. For some δ>0\delta> 0 and every t=1,…,Tt=1,\ldots,T, E(εt)=0\mathbb{E}(\boldsymbol{\varepsilon}_t)=0, E(εtεt⊤)=Φ0\mathbb{E}(\boldsymbol{\varepsilon}_t\boldsymbol{\varepsilon}_t^{\top})=\boldsymbol{\Phi}^{0}, and max⁡1≤i≤NE∣εit∣4+δ≤Cε\max_{1\le i\le N}\mathbb{E}|\varepsilon_{it}|^{4+\delta}\le C_{\varepsilon}, where τ0=(log⁡{(σ10)2},…,log⁡{(σN0)2})⊤\boldsymbol{\tau}^{0}=(\log\{(\sigma_1^{0})^2\},\ldots,\log\{(\sigma_N^{0})^2\})^{\top} and Φ0=Φ(τ0)=diag⁡{(σ10)2,…,(σN0)2}\boldsymbol{\Phi}^{0}=\boldsymbol{\Phi}(\boldsymbol{\tau}^{0})=\operatorname{diag}\{(\sigma_1^{0})^2,\ldots,(\sigma_N^{0})^2\}.

Condition 7. Let Q¨η(η0)=∂2Q(η)∂ην∂ην⊤∣η=η0\ddot{Q}_{\eta}(\eta^{0})=\left.\frac{\partial^{2}Q(\eta)}{\partial\eta_{\nu}\partial\eta_{\nu}^{\top}}\right|_{\eta=\eta^{0}} denote the population Hessian with respect to the vectorized coefficient and log-variance parameters. There exist constants 0<cH<CH<∞0<c_H<C_H<\infty, independent of NN and TT, such that,

cH≤λmin⁡{Dη,N1/2Q¨η(η0)Dη,N1/2}≤λmax⁡{Dη,N1/2Q¨η(η0)Dη,N1/2}≤CH.(10)c_H \le\lambda_{\min}\left\{\mathbf{D}_{\eta,N}^{1/2}\ddot{Q}_{\eta}(\eta^{0})\mathbf{D}_{\eta,N}^{1/2}\right\} \le\lambda_{\max}\left\{\mathbf{D}_{\eta,N}^{1/2}\ddot{Q}_{\eta}(\eta^{0})\mathbf{D}_{\eta,N}^{1/2}\right\} \le C_H. \tag*{(10)}

Condition 3 ensures that the simultaneous system remains uniformly nonsingular and that the associated spatial multipliers are uniformly bounded. Such stability restrictions are standard in the asymptotic analysis of spatial QMLE ([32, 33]).

Condition 4 ensures a regular interior parameter and stable SVD normalization. Condition 5 prevents any local block from becoming asymptotically negligible relative to the full cross section. Condition 6 assumes independence over time and across units for the errors. Its moment and independence assumptions control the linear and quadratic forms within each period and support a Lindeberg–Feller central limit theorem. Condition 7 requires the scaled population Hessian to be uniformly positive definite and bounded. It therefore provides local identification and ensures the uniform invertibility needed for the quadratic expansion of the quasi-likelihood. Related curvature conditions are commonly imposed in the inferential theory for high-dimensional generalized factor models ([39]).

We now establish average consistency for the separately identified global and local coefficient components.

Theorem 1. Suppose that Conditions 1, 3, 5, 6, and 7 hold, and let N,T→∞N,T\to\infty with N=o(T)N=o(T). Then, with probability approaching one, QTQ_T admits a local minimizer $\widehat{\eta}\in

Asymptotic Normality

HN(rg,rl)\mathcal{H}_N(r_g,r_l) satisfying

∥η^−η0∥N2=Op(T−1).(11)\lVert\widehat{\eta}-\eta^0\rVert_N^2=O_p(T^{-1}). \tag*{(11)}

Theorem 1 implies ∥A^g−Ag0∥F2=Op(T−1)\lVert\widehat{A}_g-A_g^0\rVert_F^2=O_p(T^{-1}) and ∥A^l,s−Al,s0∥F2=Op(T−1)\lVert\widehat{A}_{l,s}-A_{l,s}^0\rVert_F^2=O_p(T^{-1}) for every s=1,…,Ss=1,\ldots,S, while N−1∥τ^−τ0∥22=Op(T−1)N^{-1}\lVert\widehat{\tau}-\tau^0\rVert_2^2=O_p(T^{-1}). Thus, the global and each local coefficient matrix converge at the root-TT rate in Frobenius norm, and the average squared error of the log-variance estimates is of order T−1T^{-1}.

We next transfer the coefficient-space convergence rates to the normalized factor representation. For this purpose, write ϑ=(θ,τ)=(B,Cg,{Us,Cl,s}s=1S,τ)\vartheta=(\theta,\tau)=(\mathbf{B},C_g,\{\mathbf{U}_s,C_{l,s}\}_{s=1}^S,\tau), and let ϑv\vartheta_v denote the vector obtained by stacking these blocks in the displayed order, with each matrix vectorized columnwise. We use ϑ0\vartheta^0 and ϑ^\widehat{\vartheta} to denote the normalized true and estimated factor representations, respectively, while a generic ϑ\vartheta is not required to satisfy the normalization conditions.

To account for the different dimensions of the loading and characteristic-coefficient blocks, define the scaled factor norm by

∥ϑ−ϑ0∥F,N2=N−1∥B−B0∥F2+∥Cg−Cg0∥F2+∑s=1S{Ns−1∥Us−Us0∥F2+∥Cl,s−Cl,s0∥F2}+N−1∥τ−τ0∥22.(12)\lVert\vartheta-\vartheta^0\rVert_{F,N}^2=N^{-1}\lVert\mathbf{B}-\mathbf{B}^0\rVert_F^2+\lVert C_g-C_g^0\rVert_F^2 +\sum_{s=1}^S\left\{N_s^{-1}\lVert\mathbf{U}_s-\mathbf{U}_s^0\rVert_F^2+\lVert C_{l,s}-C_{l,s}^0\rVert_F^2\right\}+N^{-1}\lVert\tau-\tau^0\rVert_2^2. \tag*{(12)}

Corollary 1. Suppose that Conditions 1–7 hold and N=o(T)N=o(T). Let ϑ^\widehat{\vartheta} be obtained from η^\widehat{\eta} by applying the SVD normalization in Proposition 3 to the estimated coefficient matrices while leaving τ^\widehat{\tau} unchanged. After paired sign alignment,

∥ϑ^−ϑ0∥F,N=Op(T−1/2).\lVert\widehat{\vartheta}-\vartheta^0\rVert_{F,N}=O_p(T^{-1/2}).

Corollary 1 shows that the coefficient-space convergence rate is preserved by the post-estimation SVD normalization. Specifically, the normalized global and local loading matrices, characteristic-coefficient matrices, and log-variance parameters are jointly consistent at the root-TT rate under the scaled factor norm.

Asymptotic Normality

For individual asymptotic normality, consider a prespecified loading row or characteristic-coefficient column as {N,T}→∞\{N,T\} \to\infty. Specifically, 1≤i≤N1 \le i \le N for B\mathbf{B}, 1≤i≤Ns1 \le i \le N_s for Us\mathbf{U}_s, 1≤k≤K1 \le k \le K, and 1≤s≤S1 \le s \le S. Define

ϑm={Bi⊤,m=(B,i),Cg,⋅k,m=(Cg,k),Us,i,⋅⊤,m=(U,s,i),Cl,s,⋅k,m=(Cl,s,k).(13)\vartheta_m = \begin{cases} \mathbf{B}_i^\top, & m=(B,i),\\ \mathbf{C}_{g,\cdot k}, & m=(C_g,k),\\ \mathbf{U}_{s,i,\cdot}^\top, & m=(U,s,i),\\ \mathbf{C}_{l,s,\cdot k}, & m=(C_l,s,k). \end{cases} \tag*{(13)}

Accordingly, dm=rgd_m=r_g for m=(B,i)m=(B,i) or (Cg,k)(C_g,k) and dm=rld_m=r_l for m=(U,s,i)m=(U,s,i) or (Cl,s,k)(C_l,s,k). We use ϑ^m\widehat{\vartheta}_m and ϑm0\vartheta_m^0 for the corresponding blocks of the sign-aligned local estimator in Corollary 1 and of the normalized truth.

For each fixed block mm, let Σm,N\Sigma_{m,N} denote the covariance matrix of the leading term in the first-order expansion of T(ϑ^m−ϑm0)\sqrt{T}(\widehat{\vartheta}_m-\vartheta_m^0). Its explicit form is given in Appendix S3.

Condition 8. For every fixed block mm in (13), there exists a positive-definite dm×dmd_m \times d_m matrix Σm\Sigma_m with 0<cΣ≤λmin⁡(Σm)≤λmax⁡(Σm)≤CΣ<∞0<c_\Sigma\le\lambda_{\min}(\Sigma_m) \le\lambda_{\max}(\Sigma_m) \le C_\Sigma< \infty. As N,T→∞N,T \to\infty, ∥Σm,N−Σm∥F→0\|\Sigma_{m,N}-\Sigma_m\|_F \to0 for m=(B,i)m=(B,i) or (U,s,i)(U,s,i), ∥NΣm,N−Σm∥F→0\|N\Sigma_{m,N}-\Sigma_m\|_F \to0 for m=(Cg,k)m=(C_g,k), and ∥NsΣm,N−Σm∥F→0\|N_s\Sigma_{m,N}-\Sigma_m\|_F \to0 for m=(Cl,s,k)m=(C_l,s,k).

Following the fixed-index formulation of Ouyang et al. [35], Condition 8 separately assumes that the appropriately scaled covariance of each fixed block has a uniformly positive definite limit. The scaling may differ between loading rows and characteristic-coefficient columns.

Theorem 2. Suppose that Conditions 11–88 hold. If N,T→∞N,T \to\infty with N=o(T1/2)N=o(T^{1/2}), then for every fixed block mm in (13),

Σm,N−1/2T(ϑ^m−ϑm0)⟶dN(0,Idm).\Sigma_{m,N}^{-1/2}\sqrt{T}\left(\widehat{\vartheta}_m-\vartheta_m^0\right)\mathrel{\stackrel{d}{\longrightarrow}}\mathcal{N}(0,I_{d_m}).

Theorem 2 gives the asymptotic normal distribution for each prespecified loading row or characteristic-coefficient column, standardized by its own covariance.

Remark 4. The average-consistency result only requires N=o(T)N=o(T), whereas Theorem 2 is stated under the stronger condition N=o(T1/2)N=o(T^{1/2}). The reason is that consistency only requires the estimation error to vanish under the scaled parameter norm, while fixed-block asymptotic normality further requires the remainder in the first-order linear representation to be negligible after covariance standardization. The covariance matrix Σm,N\boldsymbol{\Sigma}_{m,N} accounts for the different fluctuation scales across the loading and characteristic-coefficient blocks. Under this normalization, the Hessian-replacement remainder used in the proof is uniformly bounded by Op(N/T)O_{\mathrm{p}}(N/\sqrt{T}) over the four types of fixed blocks. Hence N=o(T1/2)N=o(T^{1/2}) is sufficient for this remainder to be op(1)o_{\mathrm{p}}(1) and yields a unified growth condition for Theorem 2.

Rank Selection

The preceding results assume that the global and local ranks are known. We now consider their joint selection using an information criterion. Let r0=(rg,rl)\boldsymbol{r}_0=(r_g,r_l) denote the true rank pair, and let K={(kg,kl)∈N2:1≤kg,kl≤kmax⁡}\mathcal{K}=\{(k_g,k_l)\in\mathbb{N}^2:1\leq k_g,k_l\leq k_{\max}\} be the candidate set, where max⁡{rg,rl}≤kmax⁡≤K\max\{r_g,r_l\}\leq k_{\max}\leq K is fixed. For each k=(kg,kl)∈K\boldsymbol{k}=(k_g,k_l)\in\mathcal{K}, let VN(k)\mathcal{V}_N(\boldsymbol{k}) consist of (θ,τ)(\boldsymbol{\theta},\boldsymbol{\tau}), where θ=(B,Cg,{Us,Cl,s}s=1S)\boldsymbol{\theta}=(\mathbf{B},\mathbf{C}_g,\{\mathbf{U}_s,\mathbf{C}_{l,s}\}_{s=1}^{S}), τ∈RN\boldsymbol{\tau}\in\mathbb{R}^N, B∈RN×kg\mathbf{B}\in\mathbb{R}^{N\times k_g}, Cg∈Rkg×K\mathbf{C}_g\in\mathbb{R}^{k_g\times K}, Us∈RNs×kl\mathbf{U}_s\in\mathbb{R}^{N_s\times k_l}, and Cl,s∈Rkl×K\mathbf{C}_{l,s}\in\mathbb{R}^{k_l\times K}. Their coefficient products and variance parameters are required to satisfy the compactness and stability restrictions in Condition 3, with (rg,rl)(r_g,r_l) replaced by (kg,kl)(k_g,k_l). Write HN(k)={η(θ,τ):(θ,τ)∈VN(k)}\mathcal{H}_N(\boldsymbol{k})=\{\boldsymbol{\eta}(\boldsymbol{\theta},\boldsymbol{\tau}):(\boldsymbol{\theta},\boldsymbol{\tau})\in\mathcal{V}_N(\boldsymbol{k})\}.

For later comparison across candidate ranks, let LˉT=2QT\bar{\mathcal{L}}_T=2Q_T and Lˉ=2Q\bar{\mathcal{L}}=2Q, and write k⪰r0\boldsymbol{k}\succeq\boldsymbol{r}_0 if kg≥rgk_g\geq r_g and kl≥rlk_l\geq r_l. For an underfitted candidate k⪰̸r0\boldsymbol{k}\not\succeq\boldsymbol{r}_0, let ηk∗\boldsymbol{\eta}_{\boldsymbol{k}}^* denote a population minimizer of Lˉ\bar{\mathcal{L}} over HN(k)\mathcal{H}_N(\boldsymbol{k}). For an exact-fit or overfitted candidate k⪰r0\boldsymbol{k}\succeq\boldsymbol{r}_0, set ηk∗=η0\boldsymbol{\eta}_{\boldsymbol{k}}^*=\boldsymbol{\eta}^0, which belongs to HN(k)\mathcal{H}_N(\boldsymbol{k}) and minimizes the population criterion.

For each k∈K\boldsymbol{k}\in\mathcal{K}, let (θ^k,τ^k)(\widehat{\boldsymbol{\theta}}_{\boldsymbol{k}},\widehat{\boldsymbol{\tau}}_{\boldsymbol{k}}) denote the rank-specific local QMLE, and set η^k=\widehat{\boldsymbol{\eta}}_{\boldsymbol{k}}= η^(θ^k,τ^k)\hat{\eta}(\hat{\theta}_k,\hat{\tau}_k). We define the information criterion by

IC⁡(k)=2QT,F(θ^k,τ^k)+(kg+kl)g(N,T),\operatorname{IC}(k) = 2Q_{T,F}(\hat{\theta}_k,\hat{\tau}_k) + (k_g+k_l)g(N,T),
r^∈arg⁡min⁡k∈KIC⁡(k).(14)\hat{r} \in\arg\min_{k\in\mathcal{K}} \operatorname{IC}(k). \tag*{(14)}

Algorithm 2 implements rank selection using the numerical estimates from Algorithm 1.

For exact-fit and overfitted candidates, use the local minimizer near η0\eta^0 constructed in Appendix S4.

Condition 9. There is a constant cmin⁡>0c_{\min}>0, independent of k,N,T\mathbf{k},N,T, such that every underfitted candidate k⪰̸r0\mathbf{k}\not\succeq\mathbf{r}_0 satisfies

Lˉ(ηk∗)−Lˉ(η0)≥cmin⁡N.\bar{L}(\eta_k^*)-\bar{L}(\eta^0)\geq\frac{c_{\min}}{N}.

Condition 9 requires the minimum population criterion under every underfitted rank specification to exceed Lˉ(η0)\bar{L}(\eta^0) by at least cmin⁡/Nc_{\min}/N. The minimization is joint over the coefficient matrices and the log-variance vector in HN(k)\mathcal{H}_N(\mathbf{k}), with ηk∗\eta_k^* denoting a population minimizer. For exact-fit and overfitted candidates, local quadratic growth around η0\eta^0 follows from Condition 7.

Remark 5 (Verification of Condition 9). Proposition 5 in Section S4.2 of the supplementary material gives a sufficient condition for Condition 9. It requires the covariance matrices at each underfitted population minimizer to remain separated from the true covariance matrices, as measured by their time-averaged squared Frobenius distance. Under the stability bounds, this gives the required population loss of order N−1N^{-1}. Section S5.6 provides numerical evidence for this loss in the designs considered.

Theorem 3. Let r^\hat{r} be defined by (14). Suppose that Condition 1, Conditions 3–7, and Condition 9 hold. If (N,T)→∞(N,T)\to\infty, N=o(T)N=o(T), Tg(N,T)→∞Tg(N,T)\to\infty, and Ng(N,T)→0Ng(N,T)\to0, then Pr⁡(r^=r0)→1\Pr(\hat{r}=r_0)\to1.

The restrictions Tg(N,T)→∞Tg(N,T)\to\infty and Ng(N,T)→0Ng(N,T)\to0 balance the effects of overfitting and underfitting. Specifically, Tg(N,T)→∞Tg(N,T)\to\infty ensures that the rank penalty dominates the Op(T−1)O_p(T^{-1}) reduction in the sample criterion due to overfitting, whereas Ng(N,T)→0Ng(N,T) \to0 makes the penalty asymptotically negligible relative to the order-N−1N^{-1} population loss from underfitting. We set g(N,T)=cglog⁡(NT)/NTg(N,T) = c_g \log(NT)/\sqrt{NT}, which satisfies these restrictions when Nlog⁡2(NT)/T→0N\log^{2}(NT)/T \to0. In the simulation study, we use cg=0.05c_g = 0.05 as the baseline constant, which gives reasonable rank-selection accuracy in the designs considered. To assess robustness to this choice, we also conduct a sensitivity analysis over a range of constants in Appendix S5.5.

Algorithm 2: Rank selection: minimization of IC⁡(kg,kl)\operatorname{IC}(k_g,k_l)

Input: Maximum rank: kmax⁡k_{\max}.

Output: r^g,r^l\hat{r}_g,\hat{r}_l.

  1. for kg=1,…,kmax⁡k_g = 1,\ldots,k_{\max} and kl=1,…,kmax⁡k_l = 1,\ldots,k_{\max} do

  2. Run Algorithm 1 with ranks (kg,kl)(k_g,k_l) and calculate IC⁡(kg,kl)\operatorname{IC}(k_g,k_l);

  3. end

  4. Select r^g\hat{r}_g and r^l\hat{r}_l that minimize IC⁡(kg,kl)\operatorname{IC}(k_g,k_l) in (14);

  5. return r^g,r^l\hat{r}_g,\hat{r}_l;

Simulation Studies

We use Monte Carlo simulations to evaluate the finite-sample performance of QMLE and the accuracy of global and local rank selection. Two distinct simulation environments are considered. The first model adopts independent, idiosyncratic errors with various tail behaviors, whereas the second incorporates cross-sectional dependence and heteroskedasticity commonly observed in economic and financial applications.

Two Simulation Settings

We construct the data-generating processes based on the model specified in (5). To obtain coefficient matrices of the specified ranks and a stable simultaneous system, we generate the loading matrices and errors as follows.

Case 1. We first generate matrices B0\mathbf{B}_0 and Cg0\mathbf{C}_{g0}. Let QB∈RN×rg\mathbf{Q}_B \in\mathbb{R}^{N \times r_g} and QC,g∈RK×rg\mathbf{Q}_{C,g} \in\mathbb{R}^{K \times r_g} be orthonormal bases obtained from independent standard Gaussian matrices. We consider rg,rl∈{1,2}r_g,r_l \in\{1,2\}. For rg=1r_g = 1, let Dg\mathbf{D}_g be the 1×11 \times1 matrix with entry 1; for rg=2r_g = 2, let Dg=diag⁡(1,0.5)\mathbf{D}_g=\operatorname{diag}(1,0.5). We construct B0=QBDg1/2\mathbf{B}_0=\mathbf{Q}_B\mathbf{D}_g^{1/2} and Cg0=Dg1/2QC,g⊤\mathbf{C}_{g0}=\mathbf{D}_g^{1/2}\mathbf{Q}_{C,g}^{\top}. Thus, the global characteristic coefficient matrix Ag0=B0Cg0\mathbf{A}_{g0}=\mathbf{B}_0\mathbf{C}_{g0} has rank rgr_g. Similarly, for group ss, we independently generate orthonormal bases QU,s∈RNs×rl\mathbf{Q}_{U,s}\in\mathbb{R}^{N_s\times r_l} and QC,l,s∈RK×rl\mathbf{Q}_{C,l,s}\in\mathbb{R}^{K\times r_l}. Define Dl\mathbf{D}_l in the same way using rlr_l. We construct Us0=QU,sDl1/2\mathbf{U}_{s0}=\mathbf{Q}_{U,s}\mathbf{D}_l^{1/2} and Cl,s0=Dl1/2QC,l,s⊤\mathbf{C}_{l,s0}=\mathbf{D}_l^{1/2}\mathbf{Q}_{C,l,s}^{\top}, so the local characteristic coefficient matrix Al,s0=Us0Cl,s0\mathbf{A}_{l,s0}=\mathbf{U}_{s0}\mathbf{C}_{l,s0} has rank rlr_l. At each tt, we generate a characteristic matrix Zt−1∈RK×N\mathbf{Z}_{t-1}\in\mathbb{R}^{K\times N} whose entries are independent draws from U[−1,1]U[-1,1], where KK denotes the number of characteristics, with K=3K=3 in the simulations. The global component uses Zt−1\mathbf{Z}_{t-1}, and the local component for group ss uses the corresponding columns, denoted by Zs,t−1∈RK×Ns\mathbf{Z}_{s,t-1}\in\mathbb{R}^{K\times N_s}. The resulting time-varying interaction matrix is

M0t=Ag0Zt−1+bdiag⁡{Al,10Z1,t−1,…,Al,S0ZS,t−1}.\mathbf{M}_{0t}=\mathbf{A}_{g0}\mathbf{Z}_{t-1}+\operatorname{bdiag}\{\mathbf{A}_{l,10}\mathbf{Z}_{1,t-1},\ldots,\mathbf{A}_{l,S0}\mathbf{Z}_{S,t-1}\}.

To set the scale of the coefficients, we generate a separate characteristic path of length TT and multiply B0\mathbf{B}_0, Cg0\mathbf{C}_{g0}, U0\mathbf{U}_0, and Cl0\mathbf{C}_{l0} by a common constant so that max⁡t∥M0t∥2=0.8\max_t\|\mathbf{M}_{0t}\|_2=0.8 along this path, and generate Yt=(IN−M0t)−1εt\mathbf{Y}_t=(\mathbf{I}_N-\mathbf{M}_{0t})^{-1}\boldsymbol{\varepsilon}_t, t=1,…,Tt=1,\ldots,T.

For the noise term εt\boldsymbol{\varepsilon}_t, we first draw si∼iidU[0.5,1]s_i\overset{\mathrm{iid}}{\sim}U[0.5,1] and set Φ0=diag⁡(s1,…,sN)\boldsymbol{\Phi}_0=\operatorname{diag}(s_1,\ldots,s_N). To examine sensitivity to the error distribution, we generate εt=Φ01/2zt\boldsymbol{\varepsilon}_t=\boldsymbol{\Phi}_0^{1/2}\mathbf{z}_t under three specifications: (1) the elements of zt\mathbf{z}_t are independent standard normal variables; (2) the elements of zt\mathbf{z}_t are independent t(5)t(5) variables divided by 5/3\sqrt{5/3}; and (3) the elements of zt\mathbf{z}_t are independent draws from {0.3N(0,1)+0.7N(0,4)}/3.1\{0.3N(0,1)+0.7N(0,4)\}/\sqrt{3.1}. In all cases, given Φ0\boldsymbol{\Phi}_0, the errors have mean zero and covariance matrix Φ0\boldsymbol{\Phi}_0.

Case 2. The second experiment examines estimation under blockwise cross-sectional dependence and heteroskedasticity. The coefficient and characteristic processes are generated as in Case 1, but the covariance matrix Φ0\boldsymbol{\Phi}_0 is different. Following the design in Fan et al. [21], we initialize R=IN\mathbf{R}=\mathbf{I}_N and partition the first 4⌊N/4⌋4\lfloor N/4\rfloor units into ⌊N/4⌋\lfloor N/4\rfloor blocks of size four. We randomly select ⌊N0.3⌋\lfloor N^{0.3}\rfloor of these blocks and replace each selected block by the correlation matrix whose (l,k)(l,k) entry is 0.2∣l−k∣0.2^{|l-k|} for 1≤l,k≤41\le l,k\le4. Any remaining units are uncorrelated with one another and with all other units. To incorporate cross-sectional heteroskedasticity, we draw si∼iidU[1,1.5]s_i\overset{\mathrm{iid}}{\sim}U[1,1.5], define D=diag⁡(s1,…,sN)\mathbf{D}=\operatorname{diag}(s_1,\ldots,s_N), and set Φ0=D1/2RD1/2\Phi_0 = D^{1/2} R D^{1/2}. This design assesses robustness when the errors exhibit cross-sectional correlation and heteroskedasticity.

We set the number of groups to S=3S = 3 with equal group sizes N1=N2=N3=N/3N_1 = N_2 = N_3 = N/3. The true ranks for global and local factors, (rg,rl)(r_g, r_l), vary across the set {1,2}\{1, 2\}, yielding four distinct combinations. We consider a range of sample sizes with N∈{60,90,120}N \in\{60, 90, 120\} and T∈{1000,1500,2000}T \in\{1000, 1500, 2000\}, resulting in nine (N,T)(N, T) pairs. These settings are similar to the standard data sets used in economics and include combinations similar to the cross-sectional and time-series dimensions of our industry portfolio application. For each setting, we replicate the experiment 200 times, keeping the data-generating coefficient matrices and error variances fixed across replications.

For the QMLE procedure in Algorithm 1, we use the normalized objective LT,c/T\mathcal{L}_{T,c}/T and set M=3M = 3, χ=10−2\chi= 10^{-2}, Kin=100K_{\mathrm{in}} = 100, ϖ=0.5\varpi= 0.5, and ξ=10−6\xi= 10^{-6}. For each random start, we generate initial values by drawing auxiliary loading and characteristic-coefficient entries independently from N(0,1)N(0, 1) and U[−1,1]U[-1, 1], respectively. An SVD of their products, followed by a common rescaling, yields B(0)\mathbf{B}^{(0)}, Cg(0)\mathbf{C}^{(0)}_g and, for each group, Us(0)\mathbf{U}^{(0)}_s, Cl,s(0)\mathbf{C}^{(0)}_{l,s}. Among the converged estimates, we retain the one with the highest concentrated quasi-log-likelihood.

Simulation Results for QMLE Estimation

To quantify the estimation accuracy, we report the means (Mean) and standard deviations (SD) of the estimation errors across 200 replications. We use the Frobenius norm for matrices and the Euclidean norm for vectors, scaling each by its respective dimensions to facilitate comparison. Specifically, the estimation errors for the model parameters are computed as follows: ∥A^g−Ag0∥F/NK\lVert\widehat{\mathbf{A}}_g - \mathbf{A}_{g0}\rVert_F/\sqrt{NK}, ∥B^−B0∥F/Nrg\lVert\widehat{\mathbf{B}} - \mathbf{B}_0\rVert_F/\sqrt{Nr_g}, ∥C^g−Cg0∥F/rgK\lVert\widehat{\mathbf{C}}_g - \mathbf{C}_{g0}\rVert_F/\sqrt{r_gK}, ∥A^l−Al0∥F/NK\lVert\widehat{\mathbf{A}}_l - \mathbf{A}_{l0}\rVert_F/\sqrt{NK}, ∥U^−U0∥F/Nrl\lVert\widehat{\mathbf{U}} - \mathbf{U}_0\rVert_F/\sqrt{Nr_l}, ∥C^l−Cl0∥F/SrlK\lVert\widehat{\mathbf{C}}_l - \mathbf{C}_{l0}\rVert_F/\sqrt{Sr_lK}, and ∥σ^2−σ02∥2/N\lVert\widehat{\boldsymbol{\sigma}}^2 - \boldsymbol{\sigma}_0^2\rVert_2/\sqrt{N}. Here, B0\mathbf{B}_0, Cg0\mathbf{C}_{g0}, U0\mathbf{U}_0, and Cl0\mathbf{C}_{l0} denote the corresponding normalized true parameters. We also report the response-fit error {1NT∑t=1T∥(M^t−M0t)Yt∥22}1/2\left\{\frac{1}{NT}\sum_{t=1}^{T}\lVert(\widehat{\mathbf{M}}_t-\mathbf{M}_{0t})\mathbf{Y}_t\rVert_2^2\right\}^{1/2}, reported in the columns labeled Fit error. In these metrics, lower values indicate more accurate estimation.

Tables 1 and 2 summarize the simulation results for Cases 1 and 2 under normal errors, respectively. Three patterns emerge. First, for fixed NN and ranks, the mean estimation errors and response-fit errors decrease as TT increases, while their standard deviations generally decline. Second, comparing the two cases, coefficient matrix errors are similar, whereas variance estimation and response-fit errors are larger under the dependent errors in Case 2. Nevertheless, both errors decrease with TT. Third, for fixed TT and ranks, increasing NN generally reduces coefficient matrix errors, although its effects on individual factor parameter errors are less uniform.

The supplementary material examines robustness to non-Gaussian errors (Section S5.1), unequal local ranks (Section S5.2), sparse portfolio weights (Section S5.3), and larger cross-sectional dimensions (Section S5.4). Overall, the estimation errors exhibit broadly similar patterns across these settings, supporting the robustness of the estimator’s finite-sample performance.

NNTTAgA_gBCgC_gAlA_lUClC_lσ2\sigma^2Fit error
6010000.0047(0.0004)0.0442(0.0043)0.0100(0.0034)0.0081(0.0008)0.0562(0.0058)0.0202(0.0053)0.0571(0.0052)0.0521(0.0035)
15000.0038(0.0003)0.0363(0.0035)0.0073(0.0032)0.0066(0.0006)0.0465(0.0046)0.0160(0.0037)0.0467(0.0043)0.0424(0.0027)
20000.0032(0.0003)0.0314(0.0029)0.0060(0.0025)0.0057(0.0006)0.0405(0.0042)0.0133(0.0033)0.0406(0.0036)0.0365(0.0025)
9010000.0039(0.0003)0.0444(0.0036)0.0097(0.0029)0.0068(0.0006)0.0579(0.0051)0.0193(0.0045)0.0560(0.0043)0.0521(0.0030)
15000.0031(0.0003)0.0369(0.0029)0.0072(0.0024)0.0055(0.0004)0.0477(0.0038)0.0147(0.0036)0.0455(0.0036)0.0422(0.0022)
20000.0027(0.0002)0.0320(0.0025)0.0058(0.0019)0.0048(0.0004)0.0418(0.0034)0.0122(0.0031)0.0394(0.0031)0.0364(0.0019)
12010000.0032(0.0002)0.0402(0.0026)0.0081(0.0024)0.0057(0.0004)0.0524(0.0034)0.0164(0.0036)0.0565(0.0033)0.0514(0.0022)
15000.0026(0.0002)0.0332(0.0021)0.0058(0.0020)0.0046(0.0003)0.0433(0.0030)0.0121(0.0030)0.0461(0.0028)0.0416(0.0019)
20000.0022(0.0001)0.0290(0.0018)0.0046(0.0016)0.0039(0.0003)0.0377(0.0025)0.0100(0.0024)0.0399(0.0028)0.0359(0.0016)
\multicolumn{10}{c}{Panel A: rg=rl=1r_g=r_l=1}
6010000.0047(0.0005)0.0448(0.0045)0.0105(0.0036)0.0110(0.0009)0.0657(0.0058)0.0329(0.0092)0.0564(0.0053)0.0633(0.0038)
15000.0038(0.0004)0.0369(0.0036)0.0080(0.0032)0.0089(0.0006)0.0542(0.0040)0.0247(0.0059)0.0461(0.0044)0.0514(0.0029)
20000.0033(0.0003)0.0319(0.0030)0.0067(0.0027)0.0077(0.0005)0.0471(0.0034)0.0203(0.0047)0.0400(0.0037)0.0443(0.0025)
9010000.0038(0.0003)0.0420(0.0036)0.0092(0.0030)0.0089(0.0005)0.0610(0.0039)0.0286(0.0077)0.0565(0.0044)0.0634(0.0032)
15000.0031(0.0003)0.0346(0.0028)0.0070(0.0025)0.0072(0.0004)0.0503(0.0031)0.0203(0.0046)0.0458(0.0036)0.0513(0.0024)
20000.0026(0.0002)0.0299(0.0023)0.0056(0.0020)0.0062(0.0003)0.0443(0.0026)0.0170(0.0040)0.0396(0.0030)0.0443(0.0021)
12010000.0034(0.0002)0.0436(0.0028)0.0099(0.0027)0.0079(0.0004)0.0626(0.0034)0.0301(0.0071)0.0566(0.0033)0.0633(0.0023)
15000.0027(0.0002)0.0362(0.0024)0.0071(0.0023)0.0064(0.0003)0.0526(0.0030)0.0222(0.0054)0.0462(0.0029)0.0514(0.0020)
20000.0024(0.0002)0.0317(0.0020)0.0058(0.0020)0.0055(0.0003)0.0460(0.0023)0.0176(0.0040)0.0400(0.0029)0.0442(0.0016)
\multicolumn{10}{c}{Panel B: rg=1r_g=1, rl=2r_l=2}
6010000.0064(0.0004)0.0517(0.0033)0.0168(0.0056)0.0086(0.0008)0.0614(0.0058)0.0262(0.0075)0.0572(0.0053)0.0639(0.0035)
15000.0051(0.0003)0.0432(0.0029)0.0123(0.0038)0.0069(0.0007)0.0504(0.0051)0.0207(0.0058)0.0466(0.0043)0.0518(0.0028)
20000.0044(0.0003)0.0378(0.0023)0.0100(0.0031)0.0059(0.0006)0.0437(0.0044)0.0174(0.0046)0.0406(0.0037)0.0447(0.0023)
9010000.0051(0.0003)0.0458(0.0027)0.0130(0.0043)0.0068(0.0005)0.0537(0.0043)0.0184(0.0043)0.0557(0.0043)0.0629(0.0030)
15000.0041(0.0002)0.0382(0.0022)0.0092(0.0027)0.0055(0.0005)0.0443(0.0032)0.0143(0.0034)0.0453(0.0036)0.0509(0.0023)
20000.0035(0.0002)0.0335(0.0019)0.0074(0.0020)0.0047(0.0004)0.0384(0.0029)0.0121(0.0029)0.0392(0.0031)0.0439(0.0019)
12010000.0044(0.0002)0.0441(0.0021)0.0127(0.0033)0.0057(0.0004)0.0522(0.0034)0.0173(0.0041)0.0562(0.0032)0.0624(0.0024)
15000.0035(0.0002)0.0369(0.0017)0.0091(0.0021)0.0046(0.0003)0.0433(0.0028)0.0133(0.0030)0.0459(0.0029)0.0507(0.0020)
20000.0030(0.0001)0.0326(0.0014)0.0074(0.0021)0.0040(0.0003)0.0379(0.0024)0.0109(0.0026)0.0397(0.0028)0.0438(0.0017)
\multicolumn{10}{c}{Panel C: rg=2r_g=2, rl=1r_l=1}
6010000.0068(0.0005)0.0546(0.0037)0.0183(0.0058)0.0121(0.0009)0.0728(0.0065)0.0395(0.0106)0.0554(0.0050)0.0727(0.0035)
15000.0055(0.0004)0.0457(0.0033)0.0135(0.0043)0.0097(0.0007)0.0600(0.0050)0.0300(0.0081)0.0452(0.0042)0.0589(0.0029)
20000.0048(0.0003)0.0399(0.0028)0.0107(0.0035)0.0084(0.0007)0.0524(0.0043)0.0247(0.0069)0.0393(0.0038)0.0508(0.0026)
9010000.0055(0.0003)0.0503(0.0029)0.0162(0.0047)0.0097(0.0006)0.0677(0.0047)0.0347(0.0090)0.0568(0.0044)0.0737(0.0030)
15000.0044(0.0002)0.0421(0.0022)0.0116(0.0032)0.0078(0.0004)0.0555(0.0035)0.0248(0.0066)0.0461(0.0036)0.0595(0.0022)
20000.0038(0.0002)0.0367(0.0020)0.0090(0.0025)0.0067(0.0004)0.0484(0.0028)0.0199(0.0050)0.0393(0.0030)0.0512(0.0020)
12010000.0048(0.0002)0.0487(0.0022)0.0157(0.0036)0.0085(0.0004)0.0655(0.0036)0.0325(0.0081)0.0556(0.0031)0.0725(0.0021)
15000.0039(0.0002)0.0410(0.0018)0.0112(0.0025)0.0069(0.0003)0.0551(0.0030)0.0242(0.0062)0.0454(0.0028)0.0587(0.0020)
20000.0033(0.0002)0.0363(0.0017)0.0091(0.0023)0.0059(0.0003)0.0482(0.0025)0.0192(0.0048)0.0393(0.0027)0.0506(0.0017)
\multicolumn{10}{c}{Panel D: rg=rl=2r_g=r_l=2}

Table 2. Means and standard deviations (in parentheses) for Case 2 under normal errors. Notes: This table reports mean estimation errors and standard deviations (in parentheses) for Case 2 under normal errors. The table format follows Table 1.

NNTTAgA_gBCgC_gAlA_lUClC_lσ2\sigma^2Fit error
6010000.0046(0.0004)0.0438(0.0040)0.0099(0.0035)0.0081(0.0008)0.0560(0.0058)0.0201(0.0049)0.0350(0.0032)0.0404(0.0027)
15000.0038(0.0004)0.0364(0.0034)0.0078(0.0029)0.0066(0.0006)0.0463(0.0045)0.0161(0.0040)0.0286(0.0027)0.0331(0.0020)
20000.0033(0.0003)0.0316(0.0029)0.0063(0.0026)0.0057(0.0005)0.0404(0.0037)0.0135(0.0035)0.0247(0.0024)0.0285(0.0017)
9010000.0038(0.0003)0.0442(0.0036)0.0095(0.0028)0.0067(0.0005)0.0573(0.0042)0.0191(0.0046)0.0341(0.0028)0.0399(0.0022)
15000.0031(0.0002)0.0364(0.0030)0.0071(0.0021)0.0055(0.0004)0.0475(0.0034)0.0143(0.0033)0.0278(0.0022)0.0323(0.0017)
20000.0027(0.0002)0.0317(0.0025)0.0058(0.0023)0.0047(0.0004)0.0412(0.0033)0.0117(0.0027)0.0242(0.0019)0.0278(0.0015)
12010000.0032(0.0002)0.0395(0.0028)0.0082(0.0022)0.0056(0.0004)0.0516(0.0032)0.0160(0.0036)0.0345(0.0022)0.0395(0.0019)
15000.0026(0.0002)0.0330(0.0022)0.0059(0.0019)0.0045(0.0003)0.0427(0.0028)0.0122(0.0030)0.0281(0.0019)0.0321(0.0013)
20000.0022(0.0001)0.0290(0.0019)0.0047(0.0016)0.0039(0.0003)0.0375(0.0025)0.0103(0.0026)0.0242(0.0018)0.0278(0.0013)
6010000.0047(0.0005)0.0447(0.0041)0.0106(0.0043)0.0110(0.0008)0.0652(0.0051)0.0323(0.0088)0.0343(0.0033)0.0490(0.0027)
15000.0039(0.0004)0.0368(0.0035)0.0082(0.0035)0.0089(0.0006)0.0543(0.0039)0.0241(0.0068)0.0279(0.0027)0.0399(0.0020)
20000.0033(0.0003)0.0320(0.0032)0.0069(0.0030)0.0077(0.0006)0.0475(0.0037)0.0203(0.0054)0.0241(0.0023)0.0345(0.0019)
9010000.0037(0.0003)0.0417(0.0036)0.0091(0.0030)0.0088(0.0005)0.0607(0.0039)0.0284(0.0067)0.0344(0.0027)0.0488(0.0022)
15000.0030(0.0003)0.0342(0.0031)0.0069(0.0023)0.0071(0.0004)0.0506(0.0036)0.0209(0.0057)0.0279(0.0021)0.0395(0.0017)
20000.0026(0.0002)0.0298(0.0025)0.0057(0.0022)0.0061(0.0003)0.0440(0.0024)0.0168(0.0039)0.0244(0.0018)0.0340(0.0014)
12010000.0034(0.0002)0.0431(0.0030)0.0101(0.0028)0.0078(0.0004)0.0623(0.0037)0.0298(0.0079)0.0345(0.0020)0.0488(0.0018)
15000.0027(0.0002)0.0361(0.0024)0.0071(0.0023)0.0063(0.0003)0.0520(0.0029)0.0210(0.0051)0.0280(0.0019)0.0395(0.0014)
20000.0024(0.0001)0.0316(0.0020)0.0056(0.0019)0.0055(0.0003)0.0461(0.0026)0.0172(0.0049)0.0241(0.0018)0.0342(0.0012)
6010000.0063(0.0004)0.0515(0.0034)0.0162(0.0043)0.0085(0.0007)0.0612(0.0058)0.0255(0.0064)0.0350(0.0031)0.0495(0.0024)
15000.0052(0.0004)0.0432(0.0032)0.0123(0.0037)0.0069(0.0006)0.0505(0.0043)0.0204(0.0053)0.0285(0.0028)0.0402(0.0020)
20000.0044(0.0003)0.0378(0.0028)0.0100(0.0032)0.0059(0.0005)0.0436(0.0039)0.0169(0.0043)0.0246(0.0023)0.0347(0.0019)
9010000.0051(0.0003)0.0456(0.0027)0.0131(0.0039)0.0067(0.0005)0.0537(0.0038)0.0185(0.0046)0.0339(0.0027)0.0483(0.0020)
15000.0041(0.0002)0.0382(0.0021)0.0096(0.0029)0.0054(0.0004)0.0441(0.0034)0.0146(0.0036)0.0275(0.0021)0.0392(0.0017)
20000.0035(0.0002)0.0334(0.0018)0.0076(0.0024)0.0047(0.0003)0.0380(0.0028)0.0123(0.0032)0.0240(0.0019)0.0337(0.0015)
12010000.0043(0.0002)0.0435(0.0021)0.0124(0.0029)0.0057(0.0004)0.0521(0.0035)0.0173(0.0038)0.0341(0.0021)0.0479(0.0017)
15000.0035(0.0002)0.0367(0.0017)0.0091(0.0020)0.0046(0.0003)0.0428(0.0030)0.0129(0.0029)0.0277(0.0018)0.0389(0.0014)
20000.0030(0.0001)0.0324(0.0016)0.0071(0.0016)0.0039(0.0003)0.0374(0.0025)0.0105(0.0026)0.0239(0.0017)0.0336(0.0012)
6010000.0068(0.0005)0.0545(0.0037)0.0182(0.0054)0.0120(0.0008)0.0720(0.0058)0.0391(0.0100)0.0332(0.0033)0.0557(0.0026)
15000.0055(0.0004)0.0455(0.0032)0.0134(0.0048)0.0096(0.0006)0.0593(0.0047)0.0292(0.0077)0.0271(0.0028)0.0451(0.0021)
20000.0047(0.0003)0.0398(0.0029)0.0105(0.0038)0.0083(0.0006)0.0519(0.0041)0.0239(0.0063)0.0233(0.0023)0.0388(0.0018)
9010000.0055(0.0003)0.0499(0.0029)0.0160(0.0044)0.0090(0.0006)0.0664(0.0045)0.0327(0.0080)0.0349(0.0027)0.0568(0.0022)
15000.0044(0.0002)0.0419(0.0024)0.0119(0.0033)0.0077(0.0004)0.0552(0.0036)0.0236(0.0061)0.0284(0.0022)0.0461(0.0017)
20000.0038(0.0002)0.0365(0.0021)0.0091(0.0026)0.0066(0.0004)0.0479(0.0028)0.0193(0.0043)0.0247(0.0020)0.0395(0.0014)
12010000.0047(0.0002)0.0482(0.0023)0.0158(0.0036)0.0084(0.0004)0.0652(0.0039)0.0323(0.0084)0.0334(0.0020)0.0554(0.0018)
15000.0038(0.0002)0.0406(0.0021)0.0116(0.0028)0.0068(0.0004)0.0544(0.0032)0.0231(0.0062)0.0272(0.0018)0.0449(0.0014)
20000.0033(0.0002)0.0360(0.0018)0.0091(0.0022)0.0058(0.0003)0.0476(0.0027)0.0184(0.0045)0.0234(0.0017)0.0386(0.0013)

Table 1. Means and standard deviations (in parentheses) for Case 1 under normal errors. Notes: This table reports mean estimation errors and standard deviations (in parentheses) for Case 1 under normal errors. Columns report errors for the global parameters AgA_g, B, CgC_g, local parameters AlA_l, U, ClC_l, error variances σ2\sigma^2, and response-fit error (Fit error). The number of global (rgr_g) and local (rlr_l) factors was chosen from {1,2}\{1,2\}, yielding four combinations detailed in Panels A to D.

Simulation Results for Rank Selection

We then evaluate the performance of the IC proposed in (14) for determining the numbers of global and local factors. Table 3 reports the probabilities of correctly selecting each rank and both ranks jointly using the IC in Eq. (14) under normal errors. First, for most rank combinations, the joint selection probabilities approach one as TT increases for a fixed NN. Selection is more difficult when both true ranks are larger, and the improvement with TT is not always monotone. Second, larger cross-sectional dimensions can reduce selection accuracy in smaller samples. Comparing Case 2 with Case 1, however, the presence of cross-sectional dependence and heteroskedasticity does not systematically reduce selection accuracy.

NNTTrg=rl=1r_g=r_l=1rg=1, rl=2r_g=1,\ r_l=2rg=2, rl=1r_g=2,\ r_l=1rg=rl=2r_g=r_l=2
r^g\hat r_gr^l\hat r_l(r^g,r^l)(\hat r_g,\hat r_l)r^g\hat r_gr^l\hat r_l(r^g,r^l)(\hat r_g,\hat r_l)r^g\hat r_gr^l\hat r_l(r^g,r^l)(\hat r_g,\hat r_l)r^g\hat r_gr^l\hat r_l(r^g,r^l)(\hat r_g,\hat r_l)
Panel A: Case 1
6010001.000.980.980.990.990.980.950.930.900.960.910.87
15001.001.001.001.000.990.990.961.000.960.910.890.80
20001.001.001.001.001.001.000.981.000.980.980.940.92
9010000.910.860.790.881.000.881.000.890.890.970.970.94
15001.001.001.001.001.001.001.001.001.001.000.990.99
20001.001.001.001.001.001.001.001.001.001.001.001.00
12010000.810.680.540.791.000.790.940.610.570.920.990.91
15001.000.990.991.000.990.991.001.001.000.981.000.98
20001.001.001.001.000.980.981.001.001.000.951.000.95
Panel B: Case 2
6010000.990.980.970.990.980.970.890.990.880.890.940.83
15001.001.001.001.000.990.990.931.000.930.910.910.82
20001.001.001.001.000.990.990.951.000.950.920.930.85
9010000.960.880.850.941.000.940.970.870.850.990.990.98
15001.000.990.991.001.001.001.000.990.990.960.970.93
20001.001.001.001.001.001.001.001.001.000.980.980.96
12010000.810.640.550.770.970.750.960.610.580.970.990.96
15001.000.990.991.001.001.001.001.001.000.940.980.92
20001.001.001.001.000.990.991.001.001.001.000.980.98

Table 3. Estimation accuracy of r^g\hat r_g and r^l\hat r_l for Cases 1 and 2 under normal errors. Notes: This table reports the proportions of correct rank selection for Cases 1 and 2 under normal errors. The true global (rgr_g) and local (rlr_l) ranks are set to {1,2}\{1,2\}, yielding four combinations shown in the column groups. Columns headed r^g\hat r_g and r^l\hat r_l report the proportions of correctly selecting the global and local ranks, respectively, while (r^g,r^l)(\hat r_g,\hat r_l) reports the proportion of jointly selecting both ranks correctly. Panels A and B correspond to Cases 1 and 2, respectively.

Additional results under standardized t(5)t(5) and mixture-normal errors are reported in Section S5.1 of the supplementary material, while Section S5.5 examines sensitivity to the penalty constant. Overall, the simulations indicate generally high rank-selection accuracy, with finite-sample performance depending on the sample dimensions, true ranks, and penalty choice.

Empirical Applications

In this section, we apply the HGFM to returns on S&P 100 constituents from five industry groups. We first estimate the model parameters using the QMLE method. We then compare explanatory fit and investment performance with competing models and examine transaction costs and alternative empirical specifications.

Data Description

We analyze daily returns from the Center for Research in Security Prices (CRSP) for S&P 100 constituents from January 1, 2016, to December 31, 2025. After excluding stocks with insufficient return histories, we obtain a balanced panel of 98 stocks with 2,514 daily observations per stock. The stocks are assigned to five industry groups based on the Global Industry Classification Standard (GICS), with group membership held fixed throughout the sample. Table S12 in the supplementary material reports the group composition and complete stock list. We construct the global and local portfolio weights using market beta, market equity, and book-to-market, each ranked across the sample stocks and scaled to [−1,1][-1,1]. Local factors use these standardized characteristics within each industry group.

Empirical Findings

Global and Local Factor Model Performance

The information criterion in (14) selects r^g=r^l=1\hat{r}_g = \hat{r}_l = 1, giving one global factor and one local factor for each industry group. We estimate this specification by QMLE using Algorithm 1 with three random initializations. The global factor is denoted by G1G1, and the five local factors by L1L1–L5L5, following the industry order in Table S12. We compare the full HGFM with the capital asset pricing model (CAPM), the Fama–French three-factor model (FF3) (Fama and French [17]), and the Fama–French five-factor model (FF5) (Fama and French [18]). We also include equal- and value-weighted portfolios of the sample stocks, denoted by EW98 and VW98, respectively. To assess the contributions of the global and local factors, we consider two subsets of the jointly estimated factors: HGFM (global), containing G1, and HGFM (local), containing L1–L5.

Explanatory fit. We assess explanatory performance using Total R2R^2 and cross-sectional R2R^2 (CS R2R^2) [30, 24]. Total R2R^2 measures the fit of realized excess returns across assets and dates, whereas CS R2R^2 measures the fit of average excess returns across assets. Specifically,

Total R2=1−∑i=1N∑t=1T(Yite−Y^ite)2∑i=1N∑t=1T(Yite)2,CS R2=1−∑i=1N(Y‾ie−Y‾^ie)2∑i=1N(Y‾ie)2,\mathrm{Total}\ R^2 = 1 - \frac{\sum_{i=1}^{N}\sum_{t=1}^{T}(Y_{it}^{e}-\widehat{Y}_{it}^{e})^2}{\sum_{i=1}^{N}\sum_{t=1}^{T}(Y_{it}^{e})^2}, \quad\mathrm{CS}\ R^2 = 1 - \frac{\sum_{i=1}^{N}(\overline{Y}_{i}^{e}-\widehat{\overline{Y}}_{i}^{e})^2}{\sum_{i=1}^{N}(\overline{Y}_{i}^{e})^2},

where Yite=Yit−rf,tY_{it}^{e}=Y_{it}-r_{f,t}, YitY_{it} is the iith element of Yt\mathbf{Y}_t, and rf,tr_{f,t} is the risk-free return. The quantity Y^ite\widehat{Y}_{it}^{e} denotes the fitted excess return from each model. The quantities Y‾ie\overline{Y}_{i}^{e} and Y‾^ie\widehat{\overline{Y}}_{i}^{e} denote the time averages of observed and fitted excess returns, respectively.

Table 4 reports Total R2R^2 for the full stock panel and each industry group. We find that the full model has the highest Total R2R^2 among all models considered, both for the full panel and within each industry group. For the full panel, its Total R2R^2 is 0.538, compared with 0.362 for HGFM (global), 0.524 for HGFM (local), and 0.435 for FF5, the best-performing benchmark. The local factors account for much of the explanatory fit, while adding the global factor provides further gains across all five industries. These results indicate that the global and local factors together explain realized return variation more effectively than either subset alone or the benchmark models.

ModelAllI1I2I3I4I5
Panel A: HGFM
HGFM0.5380.5000.4250.4280.7080.580
HGFM (global)0.3620.3230.3050.2410.5400.362
HGFM (local)0.5240.4810.3960.4270.7010.571
Panel B: Benchmarks
CAPM0.3440.3690.3010.2040.4800.310
FF30.4250.4130.3470.2220.6340.427
FF50.4350.4200.3630.2300.6460.436
EW980.3680.3120.3060.2330.5530.391
VW980.3320.3740.2950.2070.4510.284

Table 4. Total R2R^2 for the full stock panel and industry groups. Notes: This table reports Total R2R^2 for the full stock panel (All) and the five industry groups (I1–I5). Panel A presents results for the full HGFM and its global-only and local-only subsets, using factors from the joint estimation. Panel B presents results for CAPM, FF3, FF5, and the equal- and value-weighted sample portfolios (EW98 and VW98).

Table 5 reports the corresponding results for CS R2R^2. For the full panel, HGFM achieves 0.836, exceeding all benchmark models but remaining slightly below HGFM (local). At the industry level, our model outperforms all benchmarks in technology and communication (I1), but not in the other groups. Overall, the benefits of combining global and local factors are particularly evident in explaining realized returns, while improvements in average-return fit vary across industry groups.

ModelAllI1I2I3I4I5
Panel A: HGFM
HGFM0.8360.8950.7120.6900.9190.800
HGFM (global)0.7400.6660.6970.6760.9230.841
HGFM (local)0.8390.8510.7980.6920.9160.849
Panel B: Benchmarks
CAPM0.7680.7070.7380.6600.9280.866
FF30.7720.7070.7180.7030.9390.883
FF50.7670.7070.6830.7120.9370.889
EW980.7320.6760.6910.6700.9050.785
VW980.7930.7520.7680.6900.9270.856

Table 5. CS R2R^2 for the full stock panel and industry groups. Notes: This table reports CS R2R^2 for the full stock panel (All) and the five industry groups (I1–I5). The table format follows Table 4.

Investment performance. We evaluate investment performance using mean-variance efficient (MVE) portfolios formed from the factors. For multifactor models, portfolio weights are determined using the means and covariance matrix of daily factor excess returns. Single-factor models use the factor directly. Table 6 reports investment performance before transaction costs. The full HGFM achieves the highest annualized Sharpe ratio among the models considered, at 1.32, compared with 0.84 for HGFM (global), 1.09 for HGFM (local), and 0.98 for FF3. It also has the lowest volatility and maximum drawdown, despite a lower mean return than either factor subset. These results suggest that combining global and local factors improves the balance between return and risk within the evaluation sample.

ModelAVG (%)SD (%)SRMDD (%)
Panel A: HGFM
HGFM0.751.481.328.12
HGFM (global)1.224.240.8420.20
HGFM (local)1.714.861.0928.94
Panel B: Benchmarks
CAPM1.264.540.8224.90
FF30.842.320.9812.23
FF50.671.740.979.75
EW981.324.480.8721.56
VW981.384.430.9323.64

Table 6. Investment performance. Notes: This table reports investment performance across different models. AVG and SD are the mean and standard deviation of monthly total returns (%). SR is the Sharpe ratio of monthly excess returns, annualized by 12\sqrt{12}. MDD is the maximum drawdown based on monthly returns (%).

We further examine transaction costs and alternative industry groupings in Tables S13 and S14 of the supplementary material. HGFM retains a higher Sharpe ratio than the equal- and value-weighted sample portfolios at all reported transaction cost rates. The alternative grouping provides an additional check on the specification, showing that explanatory fit and investment performance depend on how industries are grouped.

Economic Interpretation

We interpret the factor portfolios through the estimated characteristic coefficients that determine their stock weights. We report C^g\widehat{C}_g and C^l,s\widehat{C}_{l,s} after the weighted SVD normalization in Proposition 3. The entries of C^g\widehat{C}_g and C^l,s\widehat{C}_{l,s} are the estimated coefficients used to form portfolio weights, rather than coefficients from a separate regression of factors on characteristics. The weights are C^gZt−1\widehat{C}_g Z_{t-1} and C^l,sZs,t−1\widehat{C}_{l,s} Z_{s,t-1} before rescaling each portfolio so that its absolute stock weights sum to one. A positive coefficient means that, for two stocks with the same values of the other characteristics, the stock with the higher value of that characteristic receives a higher weight before this rescaling. These coefficients describe portfolio composition, not risk premia.

Table 7 reports the entries of C^g\widehat{C}_g and C^l,s\widehat{C}_{l,s} corresponding to market beta, market equity, and book-to-market. Several patterns emerge from these estimates. First, the global factor assigns greater weight to larger stocks and stocks with higher book-to-market. ratios, while its market-beta coefficient is small and negative. Second, local factors differ across industries. In technology and communication (I1), L1 favors larger, higher-beta stocks with lower book-to-market ratios, whereas L2 in the consumer group (I2) favors larger, lower-beta stocks with higher book-to-market ratios. Third, market equity has the largest coefficient for healthcare (I3), while book-to-market has the largest coefficient for financials (I4) and industrials, energy, and utilities (I5). L5 also assigns greater weight to higher-beta and smaller stocks. These patterns illustrate how observable characteristics shape the global and local portfolio weights. Section S6.4 of the supplementary material reports stock weights and gross-exposure shares for these factor portfolios.

FactorMarket betaMarket equityBook-to-market
G1-0.0140.1270.046
L10.2080.150-0.235
L2-0.3390.3200.243
L30.0650.2260.112
L40.032-0.0300.265
L50.128-0.0500.225

Table 7. Estimated characteristic coefficients of global and local factors. Notes: This table reports the estimated coefficients on market beta, market equity, and book-to-market used to construct the global and local portfolio weights. G1 denotes the global factor, and L1–L5 denote the local factors.

Conclusion

This paper proposes a Hierarchical Group Factor Model (HGFM) with global and local factors constructed from observed returns using characteristic-driven weights. Conditional on identification of the time-varying interaction matrix, cross-group characteristic variation separates the global and local coefficient matrices. Normalization using the empirical second-moment matrices of the observed characteristics identifies their factor representations up to paired sign changes. Under a fixed characteristic design and the stated conditions, we establish local consistency for the coefficient and log-variance estimators, as well as consistency and fixed-block asymptotic normality for the normalized factor parameters. We also propose an information criterion that consistently selects the global and common local ranks.

In the application to S&P 100 constituents, the global and local factor portfolios place different weights on stock characteristics. Combining these factors improves realized-return fit across all industry groups relative to either subset alone and the benchmarks, while delivering the highest Sharpe ratio and lowest volatility and maximum drawdown among the portfolios considered. The combined portfolio’s Sharpe ratio remains above those of equal- and value-weighted sample portfolios after transaction costs, although performance varies with industry grouping.

We suggest four directions for future research. First, the model can be extended to incorporate exogenous predictors and autoregressive dynamics, thereby allowing it to capture observed covariate effects and temporal dependence. Second, the framework could be extended to allow the number of characteristics to grow, subject to further identification conditions and asymptotic analysis. Third, this paper focuses on grouped vector-valued observations. Related hierarchical structures could be considered for matrix-valued time series (e.g., [38, 40]), three-level factor structures (e.g., [29]), and tensor factor models (e.g., [11, 10]). Finally, the model could be used to examine global and local components within broader collections of candidate factors ([28]) and incorporate machine-learning-based factor construction.

Data Availability Statement The data that support the findings of this study are available from Wharton Research Data Services. Restrictions apply to the availability of these data, which were used under license for this study.

References

  1. [1]Ahn, S. C. and A. R. Horenstein (2013). Eigenvalue ratio test for the number of factors. Econometrica 81(3), 1203–1227.DOI
  2. [2]Ando, T. and J. Bai (2017). Clustering huge number of financial time series: A panel data approach with high-dimensional predictors and factor structures. Journal of the American Statistical Association 112(519), 1182–1198.
  3. [3]Andreou, E., P. Gagliardini, E. Ghysels, and M. Rubin (2019). Inference in group factor models with an application to mixed-frequency data. Econometrica 87(4), 1267–1305.
  4. [4]Andreou, E., P. Gagliardini, E. Ghysels, and M. Rubin (2025). Spanning latent and observable factors. Journal of Econometrics 248, 105743.
  5. [5]Bai, J. (2003). Inferential theory for factor models of large dimensions. Econometrica 71(1), 135–171.DOI
  6. [6]Bai, J. and K. Li (2012). Statistical analysis of factor models of high dimension. Annals of Statistics 40(1), 436–465.DOI
  7. [7]Bai, J. and S. Ng (2002). Determining the number of factors in approximate factor models. Econometrica 70(1), 191–221.DOI
  8. [8]Bai, J. and S. Ng (2013). Principal components estimation and identification of static factors. Journal of Econometrics 176(1), 18–29.DOI
  9. [9]Brandt, M. W., P. Santa-Clara, and R. Valkanov (2009). Parametric portfolio policies: Exploiting characteristics in the cross-section of equity returns. Review of Financial Studies 22(9), 3411–3447.
  10. [10]Chen, E. Y., D. Xia, C. Cai, and J. Fan (2024). Semi-parametric tensor factor analysis by iteratively projected singular value decomposition. Journal of the Royal Statistical Society Series B: Statistical Methodology 86(3), 793–823.DOI
  11. [11]Chen, R., D. Yang, and C.-H. Zhang (2022). Factor models for high-dimensional tensor time series. Journal of the American Statistical Association 117(537), 94–116.
  12. [12]Choi, I., R. Lin, and Y. Shin (2023). Canonical correlation-based model selection for the multilevel factors. Journal of Econometrics 233(1), 22–44.
  13. [13]Choi, S. H. and D. Kim (2023). Large volatility matrix analysis using global and national factor models. Journal of Econometrics 235(2), 1917–1933.
  14. [14]Cong, L. W., G. Feng, J. He, and J. Li (2024). Sparse modeling under grouped heterogeneity with applications to asset pricing. Working paper, https://doi.org/10.2139/ssrn.4511953.
  15. [15]Cui, L., G. Feng, and J. Yang (2026). Heterogeneous predictability on mutual fund alphas: A sparse clustering GMM approach. Journal of Econometrics 257, 106300.DOI
  16. [16]Ding, Y., Y. Li, and X. Zheng (2021). High dimensional minimum variance portfolio estimation under statistical factor models. Journal of Econometrics 222(1), 502–515.DOI
  17. [17]Fama, E. F. and K. R. French (1993). Common risk factors in the returns on stocks and bonds. Journal of Financial Economics 33(1), 3–56.DOI
  18. [18]Fama, E. F. and K. R. French (2015). A five-factor asset pricing model. Journal of Financial Economics 116(1), 1–22.DOI
  19. [19]Fan, J., Y. Ke, and Y. Liao (2021). Augmented factor models with applications to validating market risk factors and forecasting bond risk premia. Journal of Econometrics 222(1), 269–294.
  20. [20]Fan, J. and Y. Liao (2022). Learning latent factors from diversified projections and its applications to over-estimated and weak factors. Journal of the American Statistical Association 117(538), 909–924.
  21. [21]Fan, J., Y. Liao, and J. Yao (2015). Power enhancement in high dimensional cross- sectional tests. Econometrica 83(4), 1497–1541.arxiv.org/abs/1310.3899
  22. [22]Fan, J. and H. Peng (2004). Nonconcave penalized likelihood with a diverging number of parameters. Annals of Statistics 32(3), 928–961.DOI
  23. [23]Feng, G., S. Giglio, and D. Xiu (2020). Taming the factor zoo: A test of new factors. Journal of Finance 75(3), 1327–1370.
  24. [24]Feng, G., J. He, N. G. Polson, and J. Xu (2024). Deep learning in characteristics-sorted factor models. Journal of Financial and Quantitative Analysis 59(7), 3001–3036.DOI
  25. [25]Giglio, S. and D. Xiu (2021). Asset pricing with omitted factors. Journal of Political Economy 129(7), 1947–1990.
  26. [28]Jensen, T. I., B. Kelly, and L. H. Pedersen (2023). Is there a replication crisis in finance? Journal of Finance 78(5), 2465–2518.
  27. [29]Jin, S., X. Lu, and L. Su (2025). Three-dimensional heterogeneous panel data models with multi-level interactive fixed effects. Journal of Econometrics 249, 105957.
  28. [30]Kelly, B. T., S. Pruitt, and Y. Su (2019). Characteristics are covariances: A unified model of risk and return. Journal of Financial Economics 134(3), 501–524.
  29. [31]Lam, C. and Q. Yao (2012). Factor modeling for high-dimensional time series: inference for the number of factors. Annals of Statistics 40(2), 694–726.DOI
  30. [32]Lee, L.-F. (2004). Asymptotic distributions of quasi-maximum likelihood estimators for spatial autoregressive models. Econometrica 72(6), 1899–1925.DOI
  31. [33]Lee, L.-f. and J. Yu (2010). Estimation of spatial autoregressive panel data models with fixed effects. Journal of Econometrics 154(2), 165–185.DOI
  32. [34]Liao, Y. and V. Todorov (2026). Observable versus latent factors. Journal of the American Statistical Association. Forthcoming.DOI
  33. [35]Ouyang, J., C. Cui, K. M. Tan, and G. Xu (2026). Statistical inference for covariate-adjusted and interpretable generalized latent factor model with application to testing fairness. Annals of Applied Statistics 20(1), 764–788.
  34. [37]Wan, R., Y. Li, W. Lu, and R. Song (2024). Mining the factor zoo: Estimation of latent factor models with sufficient proxies. Journal of Econometrics 239(2), 105386.arxiv.org/abs/2212.12845
  35. [38]Wang, D., X. Liu, and R. Chen (2019). Factor models for matrix-valued high-dimensional time series. Journal of Econometrics 208(1), 231–248.arxiv.org/abs/1610.01889
  36. [39]Wang, F. (2022). Maximum likelihood estimation and inference for high dimensional generalized factor models with application to factor-augmented regressions. Journal of Econometrics 229(1), 180–200.
  37. [40]Zhang, Y., Y. Hui, J. Song, and S. Zheng (2025). Multilevel matrix factor model. Journal of Econometrics 251, 106033.

Paper details

vlnc:W_qzxn4wsp
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents