jf-empirical-design

Asset-Pricing Test Design (jf-empirical-design)

When to trigger

• You have a candidate predictor / anomaly / factor and must decide how to test it
• You are unsure whether to run Fama–MacBeth, time-series factor regressions, or a panel
• You report t-stats but have not addressed the standard-error subtleties of cross-sectional asset pricing
• A referee will ask "is this data mining / does it survive multiple testing / does it work out of sample?"

Scope: this skill is for asset-pricing tests. For corporate/empirical causal effects, route to jf-identification.

Choosing the test

JF-specific standards

JF asset-pricing referees engage the JF-published canon — Sharpe (1964) CAPM, Fama–French (1992), Jegadeesh–Titman (1993) momentum, Carhart (1997) — and expect you to benchmark against the right factors (recall the FF three-factor model is JFE 1993). They also expect:

• Errors-in-variables / Shanken correction on Fama–MacBeth standard errors where betas are estimated.
• Multiple-testing discipline: a new anomaly must survive the "factor zoo" critique (Harvey, Liu & Zhu, JF) — adjusted t-thresholds, not the naive 1.96.
• Out-of-sample evidence for predictability claims, not just in-sample fit.
• Economic magnitude (Sharpe gain, alpha in bps), since JF writes for a general-interest reader.
• Exhaustive specifications go to the Internet Appendix (bundled in the same PDF; see jf-internet-appendix), keeping the body within 60 pages.

Worked vignette — a risk-vs-mispricing horse race

Illustrative numbers. A new characteristic predicts the cross-section: a long–short decile spread of 0.60%/month, raw t = 3.3. The JF question is not "is it significant?" but "is it risk or mispricing, and does it survive the canon?"

1. Benchmark against the right factors. Regress on Fama–French five factors + momentum; suppose the alpha is 0.42%/month (t = 2.9). The shrinkage from 0.60 already shows part of the spread is known-factor exposure.
2. Clear the multiple-testing bar. If mined from many candidates, t = 2.9 must beat the factor-zoo cutoff (Harvey–Liu–Zhu argue ~3.0, illustratively); a borderline t is a referee magnet.
3. Risk vs. mispricing. For risk, show the characteristic loads on priced covariance (GRS test). For mispricing, show the alpha concentrates among hard-to-arbitrage names and decays after ~6 months. JF rewards a paper that adjudicates.
4. Out-of-sample. Campbell–Thompson out-of-sample R² for the predictive version; in-sample-only reads as data-snooping.
5. Economic magnitude. Translate to an annualized Sharpe gain (≈ 0.35) so the general-interest reader sees the stake.

The full grid — all factor models, subperiods, cost nets — goes to the Internet Appendix; the body carries the alpha table, GRS test, and OOS result.

Referee-pushback patterns and the JF-specific fix

Calibration anchors for JF asset pricing

• The flagship rewards tests that adjudicate a first-order question (covariance risk vs. correctable mispricing) over a bare predictive regression.
• SE and multiple-testing conventions evolve; the EIV/Shanken correction and factor-zoo threshold are durable, but confirm the favored estimators and benchmark factor sets against recent JF issues.
• Economic magnitude in interpretable units (Sharpe gain, bps alpha) is non-negotiable in the body, since JF writes for a non-specialist reader.

Execution bridge (StatsPAI / Stata MCP)

Run the asset-pricing battery, don't just specify it. Full map: shared-resources/empirical-methods/execution-with-mcp.md. JF asset-pricing instantiation:

• Factor regressions / time-series alphas: feols with the right SEs (vcov="HC3" or Newey–West / clustered) — read the alpha and t off the return, not off a memory of the spec.
• The factor-zoo haircut (the JF-salient one): after disclosing how many signals were screened, apply romano_wolf (step-down, accounts for cross-signal correlation) or benjamini_hochberg, and report the alpha that survives the adjusted threshold — the executed Harvey–Liu–Zhu discipline.
• Fama–MacBeth + Shanken EIV correction are Stata-canonical: run them through the Stata MCP (mcp__stata-mcp__stata_do) with the vendored resources/code/ skeleton (asreg/xtfmb, Shanken-adjusted SEs) and reconcile to the Python alpha.
• Emit JF-format exhibits with etable; hand formatting to jf-tables-figures.

Report the economic magnitude (bps/month alpha, Sharpe gain) the body requires; the full factor-model grid and all screened signals go to the bundled Internet Appendix. If a server is not connected, adapt resources/code/ and flag any unverified number.

Checklist

• Test matched to the question (FM / time-series / panel)
• Standard errors correct for the design (Shanken, NW, clustering)
• New factor/anomaly survives a multiple-testing-adjusted threshold
• Out-of-sample check for any predictability claim
• Benchmarked against the standard factor models, attributed correctly
• Economic magnitude reported, not just t-stats

Anti-patterns

• Reporting raw t > 1.96 as decisive after mining many signals (the factor-zoo trap)
• Fama–MacBeth t-stats with no EIV/Shanken adjustment
• In-sample-only predictability dressed up as a discovery
• Crowding every robustness table into the body instead of the Internet Appendix

Output format

【Test chosen + why】...
【SE correction (Shanken/NW/cluster)】...
【Multiple-testing threshold cleared?】yes / no
【Out-of-sample evidence?】yes / no
【Economic magnitude】...
【Next step】jf-robustness