Survivorship Bias in Fund Data: Why Dead Funds Matter

What survivorship bias actually is

Survivorship bias is a sampling error: you study the population that made it through a filter, then quietly generalize to the population that entered the filter. The famous wartime illustration is the bomber analysis, where engineers were told to armor the spots where returning planes showed bullet holes, and the statistician Abraham Wald pointed out the opposite: armor where the holes are not, because planes hit there never came back. The missing planes were the data.

Fund databases have exactly this shape. Funds that perform badly tend to shut down, return capital, or get absorbed into other managers. Once that happens they stop generating new records, and any dataset assembled from currently active entities loses them entirely. The funds you can still observe are, by construction, the ones that did not blow up. The average performance of survivors is not the average performance of participants, and the gap is not random noise: it points in one flattering direction.

What makes this bias dangerous rather than merely annoying is that it is invisible from inside the dataset. A survivorship-biased table looks complete. Every row is real, every filing is genuine, every number reconciles to a real SEC document. Nothing in the data itself tells you that hundreds of other filers used to exist and are now gone. You can only detect the hole by comparing against an archive that kept the dead, which is why the fix is architectural rather than statistical.

How dead funds vanish from 13F datasets

The 13F mechanism makes disappearance silent by design. An institutional manager must file once it exceeds $100 million in covered US equities, quarterly, due 45 days after quarter end (in 2026: February 17, May 15, August 14, and November 16; see the 2026 deadline calendar). But there is no tombstone form. When a fund winds down, gets acquired, or falls below the threshold, the filings just stop. The SEC's own 13F FAQ describes when the obligation begins and ends; nothing in the process announces an ending to data consumers.

So a fund can exit your dataset through several doors, all of them quiet:

Note the third row carefully. The threshold creates a censoring effect on the way down: a fund that shrinks from large to small drops out of the record during its worst stretch, which is precisely the stretch a backtest needs.

In our Q1 2026 dataset, 1,824 institutional filers reported 1.87 million long positions worth $53.7 trillion. That is one quarter's roster. The union of every manager that has filed across past years is necessarily larger than any single quarter's roster, and the difference between those two sets is the graveyard. A pipeline that only indexes the current roster has thrown the graveyard away before analysis even starts.

What survivorship bias does to smart money backtests

The standard "smart money" exercise is portfolio cloning: pick a set of respected managers, replicate their disclosed long positions each quarter, and measure the result. Survivorship bias corrupts this at the very first step, fund selection. If you choose managers from today's filer list, or from any list of names you recognize, you have conditioned on survival and on success. Managers become famous partly by surviving. You are, in effect, asking which planes came back.

The damage shows up in three predictable ways. Average returns are overstated, because the funds that would have dragged the average down exited the sample. Risk is understated, because blowups are the strongest exit mechanism of all, so the sample's drawdown history is censored. And performance persistence looks stronger than it is, because losers leaving the pool makes winners appear to repeat. Academic work on fund databases has generally found that survivorship inflates measured average performance by a meaningful margin, and the effect compounds with the length of the lookback window: the further back your backtest starts, the more dead funds it should contain, and the more flattering their absence becomes.

Survivorship rarely travels alone. Three siblings to check at the same time: backfill bias (a fund enters a database and imports its strong early history with it), look-ahead bias (using Q1 holdings before the May deadline on which they actually became public), and plain selection bias (studying only large, famous funds). The look-ahead problem is specific to 13F timing and is covered in how accurate is 13F data; consensus snapshots like the most-owned stocks of Q1 2026 are only honest because they are dated to a filing window.

How point-in-time archives fix it

The fix is not a clever statistical correction. It is keeping the data. A point-in-time archive is append-only: every filing is stored as filed, stamped with its SEC EDGAR accession number, and never deleted or overwritten when the filer later dies, merges, or amends. Amendments layer on top of originals instead of replacing them. The question "who held what as of Q2 2023" is answered from the archive of filings that existed then, not from today's roster joined backwards.

This is how Arkolith's dataset is built, and it is the same provenance discipline that keeps AI agents from hallucinating market data: every datapoint traces to a specific accession number you can verify on EDGAR full-text search. A dead fund's final filings remain queryable forever, sitting beside long-lived filers like Berkshire Hathaway whose history spans the whole archive.

Enumerating filers is one API call:

curl -H "Authorization: Bearer YOUR_KEY" "https://arkolith.com/api/v1/funds"

And pulling a specific manager's holdings history by CIK, including managers that no longer file, is another:

curl -H "Authorization: Bearer YOUR_KEY" "https://arkolith.com/api/v1/funds/1067983/holdings"

The same archive backs the human-facing investor leaderboard, and the quickstart covers minting a key and connecting over MCP if your consumer is an agent rather than a script.

How to test your own dataset for survivorship bias

You do not have to trust a vendor's claim of point-in-time integrity. Four tests, each a few minutes:

First, plot filer counts per quarter across your full history. A survivorship-biased dataset shows counts that mostly grow toward the present, because old quarters only contain managers that still exist today. A real archive shows counts that rise and fall with the actual filing population.

Second, pick a manager you know wound down and look for their final quarters. If a famous closed fund has no history in the dataset, the graveyard was discarded.

Third, diff an old quarter's roster against the current one. The names present then and absent now are the dead; if that set is implausibly small for a multi-year gap, be suspicious.

Fourth, check amendment handling. If a 13F/A silently replaced the original it amended, the dataset is destructive, and destructive pipelines usually discard dead filers too. Searching for a specific manager is a quick way to start spot-checking:

curl -H "Authorization: Bearer YOUR_KEY" "https://arkolith.com/api/v1/search?q=tiger"

The same discipline applies beyond 13Fs. Insider filings (Form 4, due within 2 business days of the trade) and activist stakes (Schedule 13D, due within 5 business days) have their own disappearance modes when issuers delist or insiders depart, and the 51,000+ insider transactions we track are kept under the same append-only rule, with provenance attached to every response.

Frequently asked questions about survivorship bias

What is survivorship bias in simple terms?

It is the error of studying only the things that made it. If you measure the average performance of funds that still exist, you ignore every fund that failed and closed, so your average is biased upward. The missing entities are the most informative ones, and they are missing precisely because of the outcome you are trying to measure.

Does SEC EDGAR itself have survivorship bias?

No. EDGAR retains every filing ever made, including the final 13Fs of funds that later closed, so the primary record is complete. The bias is introduced downstream, when an aggregator builds its database from currently active filers or overwrites history during updates. That is why per-datapoint provenance back to accession numbers matters: it lets you audit the pipeline against the unbiased source.

How much does survivorship bias inflate backtest returns?

There is no single universal number; it depends on the asset class, the database, and the length of the sample. Academic work on fund databases has generally found a meaningful upward distortion in average reported performance, large enough to flip marginal strategies from attractive to unremarkable. The longer the lookback, the larger the effect, because more of the true sample has died.

How do I avoid survivorship bias when cloning hedge fund portfolios?

Select managers using only information that was available at each historical date, from the full roster of filers at that date, including ones that later disappeared. Run the backtest on as-filed, point-in-time holdings with the 45-day disclosure lag respected. In practice that requires an append-only archive; you cannot reconstruct the graveyard from a dataset that already buried it.

This article explains public filings and data concepts. It is not investment advice.