Which Portfolios? The Construction Dependence of Factor Model Performance

Battle Count: 65

The paper is highly relevant to quantitative trading in terms of understanding how factor model evaluation depends on portfolio construction choices. For practitioners, the findings that buy-and-hold favors FF5/FF6 while daily constant-weighting favors FF3 directly impact how one should interpret factor exposures and pricing errors in different portfolio management contexts. The distinction between factor spanning (maximum-Sharpe criterion) and portfolio-level absolute fit is crucial for strategy design. However, the paper is primarily an academic evaluation methodology paper rather than a direct trading strategy paper. It does not propose implementable strategies or address transaction costs. The insights are most valuable for quantitative researchers designing factor-based strategies and for those selecting benchmark models for risk attribution.

The methodology involves constructing 500 random portfolios across 8 baseline designs (2 sampling rules x 4 weighting schemes), multiple selection ratios (9 levels), 5 rebalancing frequencies, and 3 different investable universes. Each portfolio requires daily return tracking, delisting treatment, and factor regression. The statistical tests include GRS joint-intercept tests, HAC Wald tests, Newey-West standard errors, and block-bootstrap inference. While the individual components are standard in asset pricing research, the full experimental design with 64+ cells per analysis and multiple robustness checks requires substantial computational resources and careful implementation. The paper provides detailed equations and procedures, making replication feasible for experienced quantitative researchers.

The paper uses a fixed random seed (20260614) for all stochastic procedures. Data sources are well-documented: CRSP daily stock data (Jan 1967 - Dec 2024), Kenneth R. French Data Library for Fama-French factors and portfolios, and global-q for q5 factors. The screening procedure, portfolio construction rules, and statistical tests are fully specified with equations. However, the full code for generating 500 portfolios across 8 designs, multiple selection ratios, and 5 rebalancing frequencies is not explicitly provided as a repository link. The methodology is transparent and detailed enough for replication by a skilled researcher.