Use this skill to make the research design defensible for JFQA, an empirical and quantitative finance journal. JFQA referees press hard on whether a correlation is causal (or, in asset pricing, whether a premium is robust and not data-mined).
Pick the design that matches the question and defend it:
JFQA also publishes theory. If your paper is a model, pivot this skill to: stating assumptions transparently, deriving results/propositions, clean proof exposition, and testable implications a finance reader can take to data. Keep generality matched to the question.
| Endogeneity threat | How it surfaces in the draft | JFQA-grade remedy |
|---|---|---|
| Reverse causality | outcome plausibly drives the regressor | timing structure, a shock that moves only the regressor, or an IV with an economic exclusion story |
| Omitted firm-level variation | "we control for size and B/M" | firm FE plus a within-firm variation count showing the coefficient is still identified |
| Selection into treatment | treated and control firms differ pre-event | matching or entropy balancing plus pre-trend evidence, not either alone |
| Anticipation of regulation | effects appear before adoption | shift the event date, drop the anticipation window, show announcement-date returns |
| Data-mined anomaly | one sort, one sample, large t-stat | sub-period splits, out-of-sample evidence, multiple-testing discussion |
| Bad controls | post-treatment variables on the RHS | re-specify; report with and without, and explain which is the estimand |
Suppose 23 states adopt a disclosure rule between 2008 and 2016 and the outcome is the credit spread of in-state issuers. A naive TWFE regression gives -4.1%; the Callaway-Sant'Anna group-time ATT gives -2.6% because late-vs-already-treated comparisons inflated the TWFE number. The JFQA presentation: CS estimator as the headline, TWFE relegated to the appendix with the discrepancy explained, an event-study figure whose lead coefficients are jointly insignificant (p = 0.42), and — with only 23 clusters — wild cluster bootstrap inference (p = 0.03) instead of leaning on asymptotics. That package answers the three referee questions (estimator, pre-trends, inference) before they are asked.
Estimate and audit the identification claim, don't only argue it. Full map:
execution-with-mcp. JFQA is empirical finance (asset pricing + corporate) — the DiD / IV / RDD chain for corporate causal claims, the factor-zoo haircut for cross-sectional pricing.
detect_design → recommend → fit with as_handle=true → audit_result to list
the checks the design still owes.callaway_santanna / sun_abraham + bacon_decomposition +
honest_did_from_result (the pre-trend test is low-power, Roth 2022).effective_f_test + an anderson_rubin_ci (valid under weak instruments),
not a 2SLS t-stat alone.rdrobust (bias-corrected) + rddensity / mccrary_test for manipulation.oster_delta / sensemakr — how strong a confounder would have to be.Report the economic magnitude; route the full battery to the appendix; keep every
number reproducible. A run end-to-end (synthetic data, real returns) is in the
JF execution walkthrough. If StatsPAI/Stata are not connected, adapt the
vendored resources/code/ skeleton and flag any unverified number.
【Design】sorts/FMB / panel FE / staggered DID / IV / RDD / event study / theory
【Identifying variation】what makes it credible
【Inference】clustering / weak-IV / multiple-testing handling
【Economic magnitude】effect size in finance units
【Next step】jfqa-data-analysis