Survivorship Bias in Financial History
The returns you see are the markets that survived — the losers were quietly erased
Survivorship bias in financial data occurs when only the winners remain in a dataset, making historical performance look better than the actual experience of investors at the time. The DMS database was built specifically to correct this: when constructing a UK back-history, earlier researchers used the companies that were still large and important in 1955, then traced them back to 1918 — systematically excluding firms that had already gone bankrupt or performed poorly. The result was a built-in outperformance that never existed.
The bias operates at three levels: individual companies excluded from index back-fills, entire markets that collapsed (Russia in 1917, China after 1949 — both valued at zero in the DMS dataset), and alternative asset classes where only the successful ones attract academic attention. Pokemon cards get studied; Pogs and Tazos do not. Lego gets a return series; Playmobil does not. The pattern is consistent: when something performs well, it becomes investable history; when it fails, it disappears.
Understanding survivorship bias reframes the standard equity-outperformance argument. The 20th century's strong returns were partly real and partly an artefact of which markets and companies got included in the record. Correcting for it — as the DMS 35-country database attempts to do — typically reduces estimated long-run returns and raises the importance of diversification across markets that might not survive.
- A historical return series is only as reliable as its inclusion of the failures, not just the survivors.
- Markets that collapsed attract no retrospective data collection — their absence flatters every benchmark that omits them.
- Alternative assets are especially prone to success bias: the collectible with a return series is the one that survived.
- Correcting for survivorship typically lowers estimated long-run returns and raises the value of global diversification.
- Data quality is a continuous process — not collecting data on losing markets is itself a form of selection.
- Identify the inclusion rule of any historical datasetBefore using any return series, ask: which assets were included and when were they included? If the back-history was constructed using today's survivors mapped backward, survivorship bias is present. Check whether the dataset explicitly includes market collapses, bankruptcies, and zero-value events.Pro tipThe DMS database is one of the few that explicitly values Russia (1917) and China (post-1949) at zero rather than excluding them.WarningTextbook data almost always uses US-only or survivor-only series — treat it as an upper bound, not a realistic expectation.
- Check for market-level survivorship, not just company-levelBeyond individual stocks, entire countries can disappear from investment history. Markets that performed poorly rarely attracted the data-collection effort of markets that did well. Any dataset that added countries because they 'seemed important' in the 1970s–1980s inherited success bias at the country level.Pro tipLook at how many countries a dataset covers across the full period — a 35-country 125-year series corrects far more bias than a 10-country 50-year one.
- Apply a success-bias discount to alternative asset return estimatesFor any alternative asset class — collectibles, hedge funds, private equity, crypto — ask which examples generated the return series. The Lego series exists; Playmobil's does not. The Penny Black stamp series exists; most failed collectibles have no series. Discount returns accordingly before comparing to liquid asset classes.Pro tipThe Bitcoin exclusion decision illustrates this: adding BTC now because it has performed well would introduce exactly the success bias the DMS framework is designed to eliminate.WarningSurvivorship and success bias compound: a dataset can select for surviving markets AND then track only surviving assets within them.
- Prefer world-index numbers over individual country numbersIndividual countries show enormous variance that tempts cherry-picking — Australia's high returns, the Netherlands' strong performance, Belgium's weaker record. The world index averages these outcomes and is the correct baseline for expected-return calibration. Only adjust from the world number with explicit, defensible reasoning.Pro tipEven the world index skews toward survivorship because it overweights successful economies; treat it as a ceiling for forward-looking assumptions.
- Re-examine any return assumption derived from a short windowThirty to forty years of data is insufficient for judging a risk premium because the window may coincide with a structural shift. UK government bond returns 1980–2020 look strong because interest rates fell from 16% — that one-directional move cannot repeat. Any projection based on that window embeds the tailwind as if it were permanent.Pro tipRequire at least 100 years of data across multiple countries before treating an asset-class return as a structural signal rather than a cyclical artefact.WarningRegulators and utilities routinely anchor allowable rates of return to short-window historical data, making this bias politically and economically consequential.
When researchers wanted a UK equity history extending back to 1918, they selected the companies that were prominent in 1955 and traced them backward, deliberately keeping constituents stable to minimize index turnover. This excluded every company that had been in the market in 1918 and subsequently failed.
Both markets had active stock exchanges before their respective revolutions. The DMS team made the deliberate decision to value all securities at zero from the point of state appropriation, rather than simply stopping coverage — the only methodologically honest choice.
A European student paid his entire three-year LBS foreign student fees by trading Beanie Babies at their peak market. Beanie Babies generated extraordinary short-term returns for early participants.
Early US stock market data series, covering the period before the standard 1926 starting point, omitted the Second Bank of the United States — a company that grew to represent roughly 30% of the entire market before becoming worthless.
Staunton and Marsh (with co-author Elroy Dimson) began building a corrected UK equity index in the mid-1980s, motivated by the known flaws of the Financial Times 30-share index. That index had been back-filled using only companies that had survived to 1955, introducing selection bias into every performance calculation that relied on it. The same problem appeared globally: most finance textbooks used US data alone, and even US data series had overlooked major failures like the Second Bank of the United States, which at one point represented ~30% of the market before becoming worthless. Dimson described the fieldwork to correct these gaps as 'financial archaeology' — physically climbing ladders in Viennese stock exchange archives to transcribe copper-plate handwritten price records for markets that no digital database had ever captured.