[Paper Review] Manifold Principle Component Analysis for Large-Dimensional Matrix Elliptical Factor Model

Zeyu Li, Yong He+2 more|arXiv (Cornell University)|Mar 26, 2022

Direction-of-Arrival Estimation Techniques4 citations

TL;DR

This paper introduces the Matrix Elliptical Factor Model (MEFM) to better capture heavy-tailed matrix-valued data, particularly in finance, and proposes Manifold Principle Component Analysis (MPCA) for robust estimation of loading spaces without moment constraints. MPCA outperforms traditional methods in finite samples, especially under heavy-tailed distributions, with MPCA_F showing superior robustness and consistency in estimating factor numbers and components.

ABSTRACT

Matrix factor model has been growing popular in scientific fields such as econometrics, which serves as a two-way dimension reduction tool for matrix sequences. In this article, we for the first time propose the matrix elliptical factor model, which can better depict the possible heavy-tailed property of matrix-valued data especially in finance. Manifold Principle Component Analysis (MPCA) is for the first time introduced to estimate the row/column loading spaces. MPCA first performs Singular Value Decomposition (SVD)for each "local" matrix observation and then averages the local estimated spaces across all observations, while the existing ones such as 2-dimensional PCA first integrates data across observations and then does eigenvalue decomposition of the sample covariance matrices. We propose two versions of MPCA algorithms to estimate the factor loading matrices robustly, without any moment constraints on the factors and the idiosyncratic errors. Theoretical convergence rates of the corresponding estimators of the factor loading matrices, factor score matrices and common components matrices are derived under mild conditions. We also propose robust estimators of the row/column factor numbers based on the eigenvalue-ratio idea, which are proven to be consistent. Numerical studies and real example on financial returns data check the flexibility of our model and the validity of our MPCA methods.

Motivation & Objective

• To address the limitation of existing matrix factor models that require fourth-order moments, which are often violated in financial data with heavy tails.
• To develop a robust estimation framework for matrix factor models that does not rely on moment conditions for factors or idiosyncratic errors.
• To propose a novel matrix elliptical factor model (MEFM) that generalizes existing models by incorporating elliptical distributions, including heavy-tailed ones like the matrix t-distribution.
• To introduce Manifold Principle Component Analysis (MPCA) as a new method for estimating row and column loading spaces via local SVD averaging on Grassmann manifolds.
• To establish consistent estimators for the number of row and column factors using eigenvalue-ratio techniques under the MEFM framework.

Proposed method

• Proposes the Matrix Elliptical Factor Model (MEFM), where the factor matrix and idiosyncratic errors follow a joint Elliptical Matrix Distribution (EMD), allowing for heavy-tailed data structures.
• Introduces Manifold Principle Component Analysis (MPCA), which performs local SVD on each matrix observation and averages the resulting subspaces across all observations on the Grassmann manifold.
• Develops two variants: MPCA_F using Frobenius norm optimization and MPCA_op using operator norm optimization for robust estimation of loading spaces.
• Derives theoretical convergence rates for estimators of loading matrices, factor scores, and common components under mild moment conditions.
• Proposes eigenvalue-ratio-based estimators for the number of row and column factors, proven to be consistent under the MEFM.
• Employs rolling-window validation and MSE/opMax metrics to compare MPCA with $(2D)^2$-PCA and Projection Estimation (PE) on real financial data.

Experimental results

Research questions

• RQ1 Can a matrix factor model be extended to accommodate heavy-tailed data through elliptical distributions without requiring fourth-order moments?
• RQ2 How can the row and column loading spaces be estimated robustly when factors and idiosyncratic errors lack finite moments?
• RQ3 Does MPCA outperform classical 2D-PCA and projection-based methods in finite samples under heavy-tailed distributions?
• RQ4 Are eigenvalue-ratio estimators for factor numbers consistent under the matrix elliptical factor model?
• RQ5 Can MPCA provide stable and accurate estimation of common components and factor scores in high-dimensional matrix data with heavy tails?

Key findings

• MPCA_F consistently outperforms $(2D)^2$-PCA and PE in terms of lower bias and dispersion in row loading estimation, especially as tail heaviness increases.
• The MPCA_F estimator achieves a mean MSE of 0.7488 and mean opMax of 0.7568 at n=15, with the lowest variability across bandwidths in the rolling validation.
• The eigenvalue-ratio estimators for factor numbers are proven to be consistent under the MEFM, enabling reliable model order selection.
• In the real data application on Fama-French portfolios, MPCA_F yields the most stable and accurate predictions, with MSE and opMax values consistently lower than competing methods.
• Theoretical convergence rates for MPCA estimators are established under mild conditions, without requiring existence of fourth moments.
• Numerical studies confirm that MPCA is robust to heavy-tailed distributions, with performance degradation significantly less than classical PCA-based methods.