brycewang-stanford Internet Appendix & Replication (jfm-internet-appendix)
When to trigger
- The main text is overflowing with secondary robustness tables that should move to an appendix
- A referee asked for a check that does not deserve main-text space but must be available
- The liquidity-measure construction and sample filters are described too thinly to be reproduced
- You are preparing the replication package and must document a proprietary or restricted data feed
- It is unclear which results are load-bearing (main text) vs. supporting (appendix)
What belongs where
The discipline is the main paper must be self-contained: every claim a referee must accept to believe the headline lives in the main text. The Internet Appendix carries depth, not load-bearing claims.
| Material | Main text | Internet Appendix |
|---|---|---|
| Headline result + its identifying design | ✓ | |
| The one or two load-bearing robustness checks | ✓ | |
| Alternative liquidity-measure replications | summary line | full tables |
| Sample-filter / period / subsample variations | mention | full grid |
| Full microstructure-measure construction & formulas | brief | complete derivation |
| Data cleaning rules (trade-quote match, filters) | summary | exhaustive codebook |
| Placebo / falsification batteries | result stated | full set |
| Proofs / model details (if theory) | key steps | complete |
Building the replication package (Elsevier / JFM context)
JFM is an Elsevier journal; confirm the current data-and-code policy and whether a deposit is required at submission vs. acceptance (检索于 2026-06；以官网为准 — 待核实). Regardless of the minimum, prepare:
- Data lineage: raw feed (TAQ / LOBSTER / proprietary) → cleaning scripts → analysis files, with a README mapping each table/figure to the script that builds it.
- Codebook: every variable, its construction (effective spread formula, impact horizon, sign-classification rule), and every filter with its rationale.
- Proprietary-data path: if the feed cannot be redistributed (e.g., a vendor or exchange feed), document the access path — vendor, license, query/sample definition — so a replicator can obtain equivalent data. Never imply data can be shared that cannot.
- Environment: software versions and a lockfile so estimates reproduce (PIN/lambda estimation is sensitive to solver settings).
- Map from appendix to claim: each appendix table states which main-text sentence it supports.
Worked self-containment test (illustrative)
Read only the main text and ask: could a referee accept the headline without ever opening the appendix? If the only evidence that the result is not a measurement artifact sits in Internet Appendix Table IA.7, the paper fails — move at least one alternative-measure column into the main text and leave the full battery in the appendix. Conversely, if the main text reproduces every sub-period split and influential-name drop, it is bloated — push those down and keep a one-line summary ("results are stable across sub-periods and to dropping the most-traded names; see IA.2-IA.4"). The main text carries the claims; the appendix carries the exhaustiveness.
Proprietary-data documentation patterns
Microstructure papers routinely use feeds that cannot be redistributed (exchange order-book feeds, a broker's order data, Nasdaq ITCH, a vendor's TAQ license). The honest pattern is to document the access path precisely: the vendor/exchange, the product name, the license type, the exact query or universe/period filter used, and the cost tier if relevant. A replicator with their own license can then reconstruct the sample. State plainly what is shareable (code, derived non-proprietary aggregates) and what is not. Do not write "data available on request" for data you have no right to share — that is a common and easily caught evasion.
Checklist
- Main text is self-contained: no headline-supporting claim lives only in the appendix
- Each appendix exhibit names the main-text claim it backs
- Full liquidity-measure construction and filter rules are in the codebook
- README maps every table/figure to its build script
- Proprietary/restricted data has a documented access path, not a false promise of sharing
- Software versions / lockfile recorded so numerically sensitive estimates reproduce
- Data-and-code policy specifics verified or marked 待核实
A working appendix table of contents
A curated Internet Appendix for a typical empirical microstructure paper reads as a deliberate outline, not a dumping ground: IA.1 full variable definitions and construction formulas (every liquidity/impact measure); IA.2 the complete sample-filter grid and how many records each screen removes; IA.3 alternative liquidity-measure replications of the headline; IA.4 sub-period and asset-subset robustness; IA.5 placebo/falsification tests; IA.6 computational details (solver, starting values, tolerances, software versions); IA.7 additional figures (intraday patterns, alternative event windows). Each section opens with one line stating which main-text claim it supports. This structure lets a referee find the check they want in seconds and signals that the authors know exactly which results are load-bearing.
The appendix-to-text feedback loop
Use the appendix build as a check on the main text. If, while curating the appendix, you find a robustness check that is essential to believing the headline, that check belongs in the main text — promote it. If you find a main-text table that no reader needs to accept the claim, demote it. Done well, this pass leaves a main text where every exhibit is load-bearing and an appendix where every exhibit is genuinely supplementary, each labeled with the claim it backs. The relabeling alone often surfaces a buried fragility the authors had not noticed.
Anti-patterns
- The appendix carrying a result the main argument depends on
- A sprawling appendix of orphan tables with no link to any claim
- "Data available on request" used to dodge documenting a real access path
- Omitting the construction formula for a non-standard liquidity measure
- No record of solver/version settings for PIN/structural estimates
- Treating the appendix as a dumping ground instead of curating it
Reproducibility for numerically fragile estimates
Microstructure papers often rely on estimates that are sensitive to implementation: PIN/VPIN (likelihood with many local optima), Kyle's lambda (regression specification and sampling frequency), realized-volatility and impact measures (sampling-frequency and noise corrections), and structural model solvers. For these, the appendix and replication package must record the exact solver, starting values, multistart strategy, convergence tolerance, and software versions — otherwise a replicator gets different numbers and reads it as a failure. A short "computational details" subsection that states these, plus a lockfile pinning package versions, pre-empts the most common reproducibility dispute. Where a derived aggregate can be shared even though the raw feed cannot, ship that aggregate so the final tables at least reproduce from an intermediate file.
Output format
【Journal】Journal of Financial Markets (JFM)
【Skill】jfm-internet-appendix
【Self-containment】main text stands alone? [Y/N]
【Appendix map】each exhibit ↔ main-text claim? [Y/N]
【Codebook】measures + filters fully documented? [Y/N]
【Data access path】proprietary feed access documented (not false sharing)? [Y/N]
【Reproducibility】versions/lockfile recorded? [Y/N]
【Policy status】data-and-code policy verified or 待核实
【Next skill】jfm-writing-style