Finman Empirical Design

Empirical Design (finman-empirical-design)

When to trigger

• The sample comes from CRSP / Compustat / a vendor feed and the screens and survivorship choices are not documented
• A key variable (leverage, payout, governance index, a return measure) has several definitions and you picked one without justification
• The panel mixes frequencies, has look-ahead bias, or merges datasets on a fragile key
• Standard errors are reported but the clustering and cross-sectional/time dependence are not justified

The FM empirical-design bar

FM publishes empirical finance across corporate, asset-pricing, and banking data, so the design layer is judged on whether a competent referee could reconstruct your sample and trust your measures. The journal's "less weight on trivial robustness" stance is a double-edged sword: it means you should not bury the paper in redundant checks, but it raises the premium on getting the primary design right the first time — the screens, the variable definitions, the merge, and the inference. FM referees in corporate finance are especially alert to silent sample screens, point-in-time vs. restated accounting data, and clustering that ignores the panel's dependence structure.

The data-layer audit

Hardening sequence

1. Build the attrition table. Start from the raw universe and report the count dropped at each screen; this single exhibit answers most sample-construction doubts.
2. Pin every key variable to a source field and a definition. State winsorization (typically 1%/99%) and why; if a variable has competing definitions, justify yours and note the alternative goes to the appendix.
3. Defend point-in-time discipline. For accounting variables, use data as it would have been known; for returns, avoid look-ahead in signal construction.
4. Justify the clustering. Cluster at the level where the shocks are correlated (firm, industry, state); use two-way (firm and time) when both dimensions have common shocks; address few-cluster with wild-cluster bootstrap.
5. Report power/economic scale. State N, the dependent-variable mean, and the standard-deviation-scaled effect so a reader sees the magnitude in context.

Execution bridge (StatsPAI / Stata MCP)

Run the asset-pricing battery, don't just specify it. Full map: execution-with-mcp. Financial Management is empirical corporate finance + asset pricing; corporate-causal chain (DiD/IV/RDD) plus the factor-zoo haircut for cross-sectional pricing.

• Factor regressions / time-series alphas: feols with the right SEs (Newey–West / clustered) — read the alpha and t off the return.
• Factor-zoo haircut: after disclosing how many signals were screened, apply romano_wolf / benjamini_hochberg and report the alpha that survives.
• Fama–MacBeth + Shanken EIV are Stata-canonical — run via mcp__stata-mcp__stata_do with the vendored resources/code/ (asreg / xtfmb).
• Exhibits: etable; hand formatting to the tables/figures skill.

Report the economic magnitude (bps/month alpha, Sharpe gain); full factor grid → appendix. JF execution walkthrough.

Checklist

• Attrition table from raw universe to estimation sample, with counts per screen
• Every key variable defined, winsorization stated, source field named
• Survivorship and look-ahead bias addressed (point-in-time accounting; clean signals)
• Merge keys and match rate reported; unmatched-firm bias discussed
• Panel frequency and balanced/unbalanced status stated; entry/exit handled
• Clustering level justified; two-way / few-cluster handled where needed
• Dependent-variable mean and N reported so magnitudes are interpretable