A dossier is a chain of evidence
A research dossier should let another analyst determine what was tested, what information was available, how portfolios were constructed and which claims the evidence supports. A polished return chart cannot answer those questions alone. Start with the hypothesis, eligible universe, decision frequency, primary evaluation objective and intended claim. State whether the work is illustrative arithmetic, historical research or observed implementation.
Keep these evidence levels separate. A calculation fixture can demonstrate that turnover arithmetic is correct. A reconstructed historical portfolio can test a documented hypothesis subject to data assumptions. A broker record can establish an execution, but does not automatically establish a complete or independently evaluated strategy. None should inherit a stronger label merely because the outputs share a report template.
Make the data lineage reproducible
Record source URLs, retrieval dates, dataset versions, file hashes, field mappings, units, security identifiers and permitted uses. Preserve historical availability and revision rules alongside the values. The French Data Library notes historical data updates and provides archives for selected series; this illustrates why a named source alone does not identify an immutable dataset. Pin the actual vintage used in a study.
For issuer fundamentals, retain filing references and transformation logic rather than only a cleaned feature table. SEC developer resources provide access paths to filings and related data, but they do not validate a researcher's joins. Include reconciliation samples for difficult records such as amendments, share-class changes and delistings. Where redistribution is restricted, provide lawful retrieval instructions and metadata instead of republishing licensed rows.
Connect the method to checks and results
Specify score definitions, missing-data rules, formation cutoffs, fill assumptions, portfolio constraints, exit policy and every cost component. Keep sample splits and trial history with the method version. A reviewer should be able to identify which decisions were frozen before evaluation and which changed afterward. Store failed checks and excluded observations, not just the final favorable report.
Use a small evidence matrix. For each claim, name its artifact, reproducible check and unresolved limitation. A universe-membership file can support eligibility history; it cannot establish borrow availability. A cash-flow reconciliation can verify corporate-action arithmetic; it cannot prove every source event is complete. An evaluation ledger can demonstrate separation of samples while leaving statistical uncertainty about the strategy's economic value.
Define what would change the conclusion
List predeclared robustness checks such as longer information delays, higher costs, alternative benchmark conventions and removal of the least liquid names. Explain whether each is a diagnostic, a rejection criterion or a new experiment. A verified negative or inconclusive result is useful. Missing data, an unimplemented check and a failed economic hypothesis are distinct outcomes and should not be collapsed into one status.
Finish with a claim boundary: what was reproduced, what remains assumed, and what would require fresh data or independent evaluation. Retain the original dossier when later revisions arrive. The example matrix is a specification for evidence, not a claim that those artifacts already exist. This publication's illustrative calculations do not constitute a completed instrument study, actual backtest, provider audit or production certification.
Inspect the example
| Research claim | Required evidence | Check | What it does not establish |
|---|---|---|---|
| Historical eligibility | Dated membership and exclusions | Reconcile counts and identifiers | Complete return or borrow history |
| Causal features | Versioned facts and availability times | As-of join and revision tests | Predictive skill |
| Return accounting | Holdings, prices and event ledger | Independent cash/share reconciliation | Executable fills |
| Separated evaluation | Frozen specification and split ledger | Inspect label intervals and trial history | Statistical significance by itself |
| Net implementation estimate | Trades and explicit cost model | Recompute charges and stress assumptions | Observed broker performance |
Sources & reproduction
Primary references support the definitions and method. Numerical examples are illustrative unless a source record is explicitly identified.