Survivorship bias: rebuild the historical company universe
A historical company study needs the companies eligible at the historical date, including those that later disappeared from today’s list. Starting with current survivors can change the question before AI summarizes a single report. Reconstruct membership first, then attach outcomes and documents without deleting failed or difficult-to-match entities.
The SEC’s EDGAR API documentation describes filing histories keyed by CIK and company metadata. These records can support identity and filing checks, but they are not a complete historical investable-universe service. Market eligibility, delistings and security coverage need separately justified evidence.
Define the historical population
Specify the date, market, security types and inclusion rule. Studying listed operating companies differs from studying every registrant with an EDGAR filing. Depositary instruments, funds, multiple share classes and predecessor entities need explicit treatment. The population should answer the research question rather than mirror the easiest downloadable list.
Build a membership table with stable research entity ID, security identity, eligibility start and end, source and unresolved interval. Keep removal reasons such as merger, delisting, liquidation or scope change separate. A company leaving the population is not equivalent to an unavailable numeric observation.
Ask AI to help propose identity links or classify documented exit reasons, but require source-backed review for each consequential match. Do not allow the model to fill the historical universe from remembered famous companies. Familiarity is not evidence of eligibility, and absent familiarity does not justify omission.
Hypothetical survivor-only average
Assume a fictional historical universe contains ten companies. Eight remain in a current database and each has a measured outcome of +10 under a stipulated comparable metric. Two exited and each has an outcome of −60. The survivor-only average is +10, while the full-universe average is −4.
This deliberately simple example shows how membership changes the result. It is not a claim about actual investment returns or a reason to assign −60 to every missing company. If exited-company outcomes are unknown, preserve unknowns and report sensitivity rather than inventing failure values.
Suppose one disappeared name was merely a renamed company still represented elsewhere. Counting it again would create the opposite problem: duplication. Historical reconstruction needs both inclusion of exits and careful identity continuity. All entities and outcomes in the example are hypothetical.
Universe reconstruction worksheet
| Field | Required treatment |
|---|---|
| Historical eligibility date | Fixed before selecting survivors |
| Research entity and security | Separate company continuity from traded instrument |
| Membership interval | Start, end and supporting evidence |
| Exit reason | Documented category, with uncertainty preserved |
| Outcome coverage | Observed, missing or non-comparable |
| Current status | Descriptive field, not an inclusion requirement |
Reconcile counts: opening members plus eligible entries minus exits should equal closing members under the chosen rule. Investigate residuals. A matching problem that removes a company should appear in a coverage log, not silently reduce the denominator.
Interpret incomplete reconstruction
Compare included and unresolved groups using characteristics known at the historical date where possible. If smaller or distressed firms are disproportionately unresolved, a clean-looking dataset may still be selectively incomplete. State the pattern even when the direction of its effect cannot be quantified.
Report both the target population and the actually observed sample. If historical membership cannot be established, narrow the claim to the surviving entities studied rather than calling it a complete historical cohort. AI can organize the evidence, but fluent summaries cannot repair a population defined with hindsight. Keep universe construction reproducible so another reviewer can challenge who was included before debating the resulting average.
FAQ
Should every missing company be treated as a failure?
No. Missingness, renaming, acquisition and economic failure are different states. Investigate identity and outcome evidence separately.
Does a current company list answer a historical cohort question?
Only if its historical membership is established for that question. Current survival should not silently become the selection rule.