High dimensional alpha test for linear factor pricing model with $L_q$-norm

1. Problem Statement

The paper addresses the statistical challenge of testing for zero pricing errors (alphas) in high-dimensional linear factor pricing models.

• Context: In asset pricing models (e.g., CAPM, APT), the null hypothesis $H_0: \alpha = 0$ implies that the proposed factors fully explain expected excess returns. Rejection indicates model misspecification.
• Challenge: Modern financial datasets often have a number of securities ($N$) that is comparable to or much larger than the time dimension ($T$). Classical tests (e.g., Gibbons-Ross-Shanken) fail in this regime due to unstable covariance estimation and inversion.
• Sparsity Issue: Existing high-dimensional methods generally fall into two categories:
  1. $L_2$-based (Sum-of-squares): Powerful under dense alternatives (many small non-zero alphas) but weak under sparse alternatives.
  2. $L_\infty$-based (Max-type): Powerful under very sparse alternatives (few large non-zero alphas) but weak under dense alternatives.
• Gap: The true sparsity level of pricing errors in financial data is unknown. A single test optimized for one regime suffers significant power loss in others. The paper aims to bridge the gap between dense and sparse regimes using a continuum of norms.

2. Methodology

A. Single $L_q$-Norm Tests

The authors propose a family of test statistics based on the $L_q$-norm of the studentized alpha estimates ($t_i$), where $q$ is a fixed even integer.

• Statistic Definition: For $a \in \{2, 4, 6\}$, the test statistic is:
  $$Q_{a,N} = \frac{1}{\sqrt{N}} \sum_{i=1}^N (t_i^a - \mu_{a,v})$$
  where $\mu_{a,v}$ are the moments of a $t$-distribution with $v = T-p-1$ degrees of freedom.
• Specific Tests:
  • $L_2$ test ($a=2$): Corresponds to the sum of squared t-statistics.
  • $L_4$ test ($a=4$): Intermediate sensitivity.
  • $L_6$ test ($a=6$): Higher sensitivity to sparse signals.
• Asymptotic Theory: Under mild assumptions (fixed factor dimension, independent component model for errors, weak cross-sectional dependence), the vector of statistics $\mathbf{Q}_N = (Q_{2,N}, Q_{4,N}, Q_{6,N})^T$ converges to a multivariate normal distribution.
• Covariance Estimation: The authors derive the asymptotic covariance matrix $\mathbf{B}_N$ and propose a consistent estimator using a thresholding strategy on the sample correlation matrix of residuals to handle high-dimensional noise.

B. Theoretical Power Analysis

The paper analyzes the non-centrality parameters of these tests under local alternatives.

• Key Finding: As the order $q$ increases, the test becomes increasingly sensitive to sparse alternatives.
  • Dense/Weak Signals: $L_2$ is optimal.
  • Moderate Sparsity: $L_4$ offers a balanced compromise.
  • Sparse/Strong Signals: $L_6$ (and $L_\infty$) dominates.
• This establishes a principled continuum where the choice of $q$ adapts to the unknown sparsity of the alternative.

C. Asymptotic Independence and Cauchy Combination

To create a robust procedure that does not require prior knowledge of sparsity, the authors combine tests.

• Independence Result: A crucial theoretical contribution is proving the asymptotic independence between the Gaussian limit of the finite-order $L_q$ statistics ($q < \infty$) and the Gumbel limit of the max-type $L_\infty$ statistic.
• Cauchy Combination Test: Leveraging this independence, they propose a Cauchy combination test:
  $$T_{C,N} = \frac{1}{4} \sum_{a \in \{2,4,6,\infty\}} \tan\left\{ \pi \left( \frac{1}{2} - p_a \right) \right\}$$
  where $p_a$ are the p-values from the individual tests.
• Advantage: The Cauchy combination allows for a simple standard Cauchy null distribution, avoiding the need for computationally expensive resampling or simulation for joint calibration. It aggregates evidence across different sparsity regimes.

3. Key Contributions

1. $L_q$-Norm Framework: Development of a general $L_q$-norm testing framework for high-dimensional alpha testing, extending beyond the binary choice of $L_2$ vs. $L_\infty$.
2. Asymptotic Independence: Proof that finite-order $L_q$ statistics and the $L_\infty$ statistic are asymptotically independent under the null, providing the theoretical foundation for combining them.
3. Adaptive Cauchy Test: Proposal of a Cauchy combination test that is robust to unknown sparsity levels, offering superior finite-sample performance compared to existing methods.
4. Practical Utility: Identification of the $L_4$ test as a particularly attractive "single-procedure" compromise for practitioners who prefer simplicity over complex combinations.

4. Results

Simulation Studies

• Size Control: The proposed tests (including $L_2, L_4, L_6, L_\infty$, and the combination tests) maintain good size control (close to the nominal 5% level) across various scenarios, including heavy-tailed errors ($t_8$ distribution), latent factor contamination, and spatial cross-sectional dependence.
• Power Performance:
  • Very Sparse: $L_\infty$ and $L_6$ achieve the highest power.
  • Moderately Sparse: The Cauchy combination (CC) and $L_4$ perform best. $L_4$ is shown to be a robust single-test alternative.
  • Dense: $L_2$ dominates, but the combination tests remain competitive.
  • Comparison: The Cauchy combination generally outperforms the minimum-p benchmark ($\min P$) in the moderate sparsity regime.

Real Data Application

• Dataset: Monthly returns of S&P 500 constituents (Jan 2005 – Jan 2024).
• Model: Testing the CAPM ($H_0: \alpha = 0$) using rolling windows.
• Findings:
  • The $L_2$ and $L_4$ tests, along with the combination methods, consistently reject the CAPM over substantial periods, suggesting systematic mispricing.
  • The $L_\infty$ test is less sensitive, implying that mispricing is not driven by a few extreme outliers but rather by a moderately sparse or diffuse pattern across the cross-section.
  • Rejections were persistent, particularly during the COVID-19 pandemic, indicating structural model failure rather than isolated events.

5. Significance

This paper provides a unified and adaptive solution to the high-dimensional alpha testing problem. By introducing the $L_q$-norm perspective and proving the independence of different norm-based statistics, the authors enable the construction of combination tests that are robust to the unknown nature of market anomalies. The proposed Cauchy combination test offers a practical, computationally efficient, and theoretically sound tool for financial econometrics, significantly improving upon existing methods that are often tailored to specific sparsity assumptions. The identification of $L_4$ as a robust standalone test also offers immediate practical value for applied researchers.