技能 数据科学 经验金融研究稳健性检验指南

经验金融研究稳健性检验指南

v20260724
rfs-robustness
本指南详细阐述了顶级经验金融论文必须构建的稳健性检验电池。内容涵盖替代模型、子样本分析、异常值处理和多重检验校正,旨在指导作者构建严谨的方法论框架,确保研究结论具备统计学上的可靠性、可重复性,并能应对高强度的学术同行评审。
获取技能
329 次下载
概览

Robustness & Multiple-Testing Discipline (rfs-robustness)

When to trigger

  • The main result holds in one specification but you have not stress-tested it
  • A new return predictor / cross-sectional anomaly is the central claim
  • You tested many candidate variables and want to report the ones that "worked"
  • Reviewers will ask "does this survive [alternative spec / subsample / period]?"
  • Results may be sensitive to outliers, windsorization, or a single event

The two robustness mandates at RFS

RFS punishes fragile results and, in cross-sectional asset pricing, undisciplined multiple testing. Build the battery proactively — a result that only the authors can reproduce in one specification is treated as no result.

Two RFS-specific mechanisms raise the bar above JF/JFE:

  • Public code release is a condition of publication. Referees and post-publication readers can and do re-run your code. A robustness claim you cannot reproduce from the released code is a liability, not a footnote — design the battery so the released scripts regenerate every check.
  • Registered Reports neutralize the multiple-testing critique by construction. Because RFS offers pre-results review (the format it pioneered in finance), pre-specifying the hypothesis and the test before seeing outcomes is the strongest possible answer to "you data-mined this." Even outside the Registered Report track, pre-specification is the RFS-preferred defense. The q-factor spanning logic of Hou, Xue, and Zhang (2015) "Digesting Anomalies" (RFS 28(3)) is a model for confronting the anomaly zoo head-on.

A. General-fragility battery (all empirical papers)

  • Alternative specifications: different FE, control sets, functional forms — show the coefficient is stable.
  • Alternative measures: re-estimate with an alternative proxy for the key variable.
  • Subsamples: split by period, size, industry, region; the sign should not flip without explanation.
  • Outliers: re-run with alternative winsorization/trimming; drop influential observations.
  • Placebo / falsification: a setting where the effect should be zero.
  • Alternative clustering: show inference is not driven by an SE choice.

B. Multiple-testing discipline (cross-sectional asset pricing especially)

  • If the claim is a new predictor or anomaly, confront the data-mining critique head-on.
  • Report and discuss multiple-testing-adjusted significance (e.g., Bonferroni / Holm, FDR control, or the higher t-hurdles argued in the asset-pricing replication literature such as Harvey–Liu–Zhu).
  • Distinguish in-sample fit from out-of-sample performance; report OOS explicitly.
  • Pre-specify the hypothesis; do not present a survivor of a large specification search as if it were a single test.
  • For factor claims, run spanning tests against established factor models and report the alpha after controls.

C. Mechanism / external validity (supporting robustness)

  • Show the mechanism, not just the reduced-form effect — heterogeneity consistent with the proposed channel strengthens credibility.
  • Triangulate with a second data source or setting when feasible.

Sequencing the battery

  • Run the fragility checks before writing the main tables — a result that moves under reasonable perturbation is not ready.
  • For asset-pricing claims, treat the out-of-sample and multiple-testing checks as primary, not optional add-ons.
  • Decide early which single check per result earns a place in the main paper; the rest go to the Internet Appendix (rfs-internet-appendix).
  • If a check weakens the result, address it in the text — do not bury it or omit it; referees will re-run the obvious ones.

Execution bridge (StatsPAI / Stata MCP)

Run the battery, don't just enumerate it. Full map: execution-with-mcp. RFS is finance top-3 (with JF, JFE) — corporate-causal chain for corporate papers, factor-zoo haircut for asset pricing.

  • Many outcomes / specifications: romano_wolf (step-down FWER) or benjamini_hochberg.
  • OVB sensitivity: oster_delta / sensemakr.
  • Inference: wild_cluster_bootstrap (few clusters), twoway_cluster / conley.
  • Re-fit off one handle: audit_result(result_id) lists missing checks + the exact suggest_function for each.
  • Exhibits: etable / did_summary_to_latex from the handle — no retyped numbers.

Decisive checks in the body, exhaustive battery in the appendix. JF execution walkthrough.

Checklist

  • Main coefficient shown stable across ≥3 alternative specifications
  • Alternative measure of the key variable tested
  • Subsample / period splits reported; sign stability explained
  • Outlier/winsorization sensitivity checked
  • Placebo or falsification test included
  • For asset-pricing claims: multiple-testing adjustment + out-of-sample test reported
  • Spanning tests against standard factor models (if a factor claim)
  • Every robustness check regenerable from the code RFS will require you to release publicly
  • Robustness tables sized for the Internet Appendix, not the main paper

Anti-patterns

  • Reporting the 3 predictors that worked out of 40 tested, with no adjustment.
  • Calling a result "robust" after one alternative control set.
  • In-sample-only predictability dressed as economically meaningful.
  • A subsample sign flip mentioned only in a footnote with no explanation.
  • Dumping 30 robustness tables into the main paper instead of the IA.

Output format

【Main result】coefficient / magnitude
【Fragility checks done】[specs, measures, subsamples, outliers, placebo]
【Multiple-testing】adjustment used + OOS result (if asset pricing)
【Surviving concerns】[...]
【Where reported】main paper vs. Internet Appendix
【Next step】rfs-tables-figures
信息
Category 数据科学
Name rfs-robustness
版本 v20260724
大小 6.15KB
更新时间 2026-07-29
语言