CompanyGraph
Why a visible record of winners can make success look more predictable than it was.
The Missing Observation Changes the Question
An investor studying today’s listed companies sees firms that survived listing, competition, financing, regulation, and time. A fund database sees funds that remained open. A strategy backtest may see securities that still have a ticker and price history. The failures may have disappeared through bankruptcy, delisting, merger, liquidation, or a database’s inclusion rule.
This is survivorship bias. It does not mean the surviving companies are uninteresting or that their characteristics are irrelevant. It means that a pattern found among survivors answers a narrower question: what do the observed survivors have in common? It does not by itself answer what separated them from the firms that had the same apparent attributes but did not survive.
Where the Distortion Enters
Delisting is a particularly important boundary. A security can disappear from a current-universe screen while its investor experienced a large loss, a forced sale, or a merger at an unfavourable time. Academic work on delisting returns shows why omitting the final return can bias estimates of individual stocks and portfolios. Shumway’s study of delisting bias is evidence about the return-data problem; it is not a guarantee that every database has the same omission.
Fund histories have a similar issue. A fund that closes after poor performance may no longer appear in a database’s current list. An index that replaces a weak constituent with a stronger one records the index’s rule-based portfolio, not the experience of holding the original names forever. The selection rule can be legitimate for the product, but it must not be confused with a random sample of all candidates.
Narrative selection is harder to audit. Books, conference talks, and investor letters are more likely to be produced by people whose businesses and reputations survived. Their practices may be valuable, but the story has already passed through a success filter. A case study should therefore distinguish the documented decisions from the claim that those decisions caused the outcome.
Why Backtests Become Too Confident
Suppose a historical screen selects companies with high margins and low leverage. If the dataset removes companies after bankruptcy, the screen is tested on a population that has already survived the worst outcome. The reported average may be inflated, the drawdown understated, and the apparent stability strengthened by the removal process itself.
The problem is not solved by adding more survivor years. A longer period can increase the number of omitted exits. Nor is it solved by saying that a failed company was not investable at the end. The investor could have owned it before the failure; the path into the failure is part of the strategy’s result.
A clean analysis records the security universe at each historical date, includes delisted and merged names under a stated return convention, and prevents future information from deciding which companies entered the past sample. Different conventions can produce different valid results, so the source and treatment should be visible.
Markets and Industries Are Also Selected
Global comparisons often focus on markets that remained open, institutions that continued to report, or industries that retained a public company. Markets disrupted by confiscation, capital controls, war, or permanent closure may be absent from the comparison. The conclusion “equities recover over long periods” can be useful for a defined surviving market, but it is not a law of every market that has existed.
Industry case studies have the same boundary. A durable retailer may share customer focus, disciplined inventory, and a strong culture with failed retailers. Those features could help, but timing, location, financing, technology, and luck may have mattered more. The survivor story is evidence for a hypothesis to test against failures, not a substitute for that test.
What the Numbers Establish
- A current index return establishes the result of the index’s inclusion and replacement rules.
- A fund database average establishes the average of the funds that remain in that database under its closure policy.
- A backtest establishes a simulated result only for its universe, delisting treatment, costs, and timing assumptions.
- A case study establishes what a surviving organization did and what happened afterward; it does not identify the missing counterexamples.
How to Repair the Inference
Write down the population before selecting the winners. Preserve failed, merged, and delisted observations; state how returns, cash distributions, and acquisition prices are handled; and compare the characteristic in survivors with the same characteristic in non-survivors. If the data cannot recover those observations, narrow the conclusion rather than filling the gap with confidence.
Survivorship bias is therefore not an argument against studying success. It is a demand to keep the denominator visible. The investor should ask who was eligible to appear, who disappeared, when they disappeared, and whether disappearance itself depended on the outcome being measured.