Specification Choice in Asset Pricing

The field's disagreements are locations in specification space, not contradictions in the data.private preview

Arman Goudarzi and Matthew Ringgenberg, University of Utah

Half a century of asset-pricing tests disagree about whether market beta is priced, how large the value and momentum premia are, and which factors matter. This project enumerates every defensible way of running the classic tests and building the canonical factors, evaluates all of them on the same data (CRSP/Compustat, 1926–2024), and places the published papers where they actually stand: each one a small, identifiable region of a much larger space. The disputes largely dissolve: different answers come from different choices, made on the same data.

929,280
CAPM test specifications
Median slope 1.15%/yr against a predicted 10.1%; positive in 78%, reliably below the premium in 86%, significant in under 2% under Newey–West.
72,576
distinct factor constructions
Every factor's mean is positive in at least 91% of its constructions, but 80% of value and 77% of momentum constructions deliver less than the published series.
2,064
full-pipeline propagation variants
Construction choices move factor means about five times more than they move premia: pass-through is roughly one-fifth. Value survives everywhere; momentum inverts everywhere.

CAPM explorer

Is market beta priced? All 929,280 answers on one curve, with exact statistics, three inference conventions, paper presets from Black–Jensen–Scholes to Welch, and point-by-point comparison of any two specifications.

Open the curve →

Factor construction explorer

HML, SMB, RMW, CMA, and UMD rebuilt every defensible way, against the published Ken French series on the same months. See exactly which construction cell each classic paper committed to.

Open the curves →

Propagation

Carry each construction variant through the entire pipeline to the cross-sectional premium. How much of a construction choice survives? About one-fifth, and which factors are “priced” depends on the layer you ask at.

Open the scatter →

How to read a specification curve

Take every choice a researcher must make before the regression runs, how betas are estimated, how portfolios are formed, which stocks are screened out, how the factor is assembled, and call each one a knob. The cartesian product of all defensible knob settings is the specification universe. Estimate the object of interest under every specification, sort the results, and plot them: that is the curve. A published paper pins some knobs and leaves the rest unstated, so a paper is a region of the curve, not a point on it. The explorers here let you stand inside any region and see what moves when a single choice changes: click two points to get the exact, complete diff of their recipes.

What the paper adds beyond displaying dispersion

Three things. First, joint inference: a count-vector moving-block bootstrap asks whether the whole curve is consistent with a null, not whether some corner of it is. Second, a literature map: the disputing papers are replicated to a stated standard and placed onto the curves, which turns fifty years of back-and-forth into geography. Third, propagation: the construction layer is connected to the premium layer inside one design, which is how we learn that construction choices are loud where factors live and quiet where premia live.

Data and reproducibility

Monthly CRSP and Compustat via WRDS, July 1926 through December 2024 (accounting data from 1951). The published Ken French series are used only as replication benchmarks; the engine reproduces them at correlations of 0.976 to 0.9986. This site serves derived per-specification statistics only, no security-level or raw licensed data. Everything on these pages is computed by the paper's notebooks; the site is a static rendering of those exact artifacts.

Citation

Goudarzi, Arman, and Matthew Ringgenberg. “Specification Choice in Asset Pricing.” Working paper, University of Utah, 2026. (Private preview: please do not circulate.)