Estimation and inference for CP tensor factor models

Bin Chen, Yuefeng Han, Qiyang YuJournal of Econometrics 2025 · OpenReview

TL;DR

This paper introduces a new method for analyzing high-dimensional, multi-dimensional data (tensors) using CP tensor factor models. Their approach provides a more robust and flexible way to estimate underlying factors and their loadings, which is crucial for understanding complex economic and financial phenomena.

What Problem Does This Paper Address?

Modern economics and finance are awash in high-dimensional, multi-dimensional data, often best represented as tensors. Traditional methods struggle with this complexity. While tensor factor models offer a promising avenue, existing approaches often rely on assumptions like orthogonal loadings or use methods that are computationally intensive or lack theoretical guarantees for inference. This paper addresses the need for a flexible, statistically sound, and computationally efficient method to estimate and infer from CP tensor factor models, particularly when factor loadings might not be orthogonal and temporal dependence is weak.

Methods

The authors propose a two-step estimation procedure:

  1. Initialization via Randomized Composite PCA (RC-PCA): This novel initialization strategy, inspired by techniques for noiseless tensor decomposition, is adapted to handle noisy CP decomposition. It uses randomized projection to accommodate closely spaced eigenvalues, making it more robust than standard PCA.
  2. Iterative Simultaneous Orthogonalization (ISO): This refinement step, based on contemporary covariance, iteratively improves the estimates of loading vectors and latent factors. It preserves the tensor structure, leading to more parsimonious models and potentially faster convergence.

The paper also introduces two generalized eigenvalue ratio-based estimators for determining the number of latent factors, extending a criterion previously used for Tucker models.

Datasets

The paper uses a dataset of characteristic-sorted portfolio excess returns. This dataset consists of monthly portfolio excess returns sorted into 10 deciles based on firm-level characteristics, spanning from January 1990 to December 2022. The final dataset used for analysis has dimensions 127x10 with a sample size of T=396.

Key Insights and Ideas

  • Contemporary Covariance is Key: Unlike methods relying on lagged autocovariances, this approach uses contemporary covariance, which is crucial for datasets with weak temporal dependence (like asset returns), broadening the applicability of CP tensor factor models.
  • Robust Initialization: The RC-PCA initialization is designed to handle challenges like closely spaced eigenvalues and noise, which are common in real-world data.
  • Flexible Factor Loadings: The CP tensor factor model, as implemented here, allows for non-orthogonal loading vectors, offering greater flexibility than some prior methods.
  • Interpretable Factors: The empirical application successfully identifies economically meaningful factors: a market factor, a long-short factor, and a volatility factor, without requiring post-hoc rotation.

Key Results

  • Theoretical Guarantees: The paper establishes a robust inferential theory, demonstrating consistency and asymptotic normality for the estimated loading vectors under relaxed assumptions.
  • Rank Selection Consistency: The proposed eigenvalue ratio estimators for the number of factors are shown to be consistent.
  • Empirical Performance: In an application to characteristic-sorted portfolios, the CP factor model outperforms benchmark approaches in out-of-sample forecasts and yields smaller cross-sectional pricing errors.
  • Simulation Studies: Extensive simulations confirm the theoretical results and demonstrate the superiority of their proposed CC-ISO method over existing approaches (AC-ISO, TPCA) in various scenarios, including non-orthogonal loadings and serial correlation in errors.

Limitations and Future Work

  • Identifiability: While the paper discusses identifiability conditions, it acknowledges that identification can be weak or lost in cases of high collinearity or unbalanced ranks/dimensions.
  • Computational Cost: While the method is generally efficient, the RC-PCA initialization can be computationally intensive, especially in scenarios with closely spaced eigenvalues.
  • Missing Values: The paper notes that handling missing values in tensor factor models is an important and challenging problem that requires substantial additional effort and is deferred to future work.
  • Extension to Other Models: The authors suggest their inferential framework could be extended to other matrix and tensor problems.
  • Alpha Testing: The paper mentions that the asymptotic normality results could be extended to construct valid t-statistics for testing portfolio alphas, a direction for future research.

Summary updated

BibTeX
@article{chen2025estimation,
  author = {Bin Chen and Yuefeng Han and Qiyang Yu},
  title = {{Estimation and inference for CP tensor factor models}},
  journal = {Journal of Econometrics},
  year = {2025},
  url = {https://openreview.net/forum?id=afsCIXLy0a},
}
Figure