Overshooting the Coordinate: Where Factor Corrections Land on Characteristic Axes

Battle Count: 50

Highly relevant for quantitative researchers and practitioners who use factor models for portfolio construction, alpha generation, and risk decomposition. The paper demonstrates that maximum-Sharpe spanning (commonly used for factor selection) does not guarantee correct pricing along the characteristic a factor was built to capture. This has direct implications for: (1) factor model selection in multi-factor strategies, (2) understanding residual alpha in characteristic-sorted portfolios, (3) evaluating whether adding a factor to a model actually prices the intended anomaly or overshoots it, (4) time-varying factor model performance assessment. The finding that HML-based models overshoot value (leaving -28 bp rank alpha) while UMD correctly flattens momentum is practically important for strategy design. However, the paper is primarily a diagnostic/evaluation tool rather than a trading strategy generator.

Moderate-to-high complexity. Requires: (1) daily CRSP/Compustat data processing with formation-date characteristic sorting, (2) construction of 201-point bridge paths with buy-and-hold wealth tracking, (3) OLS regressions at each grid point, (4) HAC covariance estimation with Newey-West, (5) positive-semidefinite adjustment of 201x201 grid covariance matrix, (6) 50,000-draw Gaussian simulation for HAC-GP critical values, (7) moving-block bootstrap with 4,999 resamples, (8) finite-sample calibration simulations. The mathematical framework (Proposition 1, FWL identity, bridge construction) is well-specified but implementation requires careful handling of boundary stocks, dividend reinvestment, and cross-grid dependence. The paper provides detailed equations and algorithmic steps.

The paper uses standard publicly available data (CRSP, Compustat, Fama-French Data Library, Global-q.org, Global Factor Data). A replication archive is mentioned containing complete machine-readable model-by-window output. The methodology is fully specified with equations, grid parameters (201 cutoffs), bootstrap settings (B=50,000 for HAC-GP, B=4,999 for block bootstrap), and formation conventions. However, no explicit GitHub repository URL is provided. The construction is mechanical and deterministic given data inputs.