Insights

What Is Survivorship Bias in Backtesting?

Alphanume Team · April 30, 2026

Why delisted names must stay in your universe — and what happens to results when they don't.

Survivorship bias is the systematic distortion that arises when a backtest uses only securities that "survived" to the present day. Delisted, acquired, or bankrupt companies are dropped, which produces artificially attractive historical performance for any strategy that would have held those names. For short-side research particularly, the effect is severe — the worst-performing names are exactly the ones most likely to delist, and excluding them obscures the gross alpha of structural-shorting strategies.

How the bias creeps in

The standard pattern: a researcher pulls the current S&P 500 (or any other index) constituents from a vendor, downloads historical price data for those tickers, and runs a backtest. The implicit assumption — that the current universe was the universe at every historical date — is almost always wrong.

Names that delisted before today are absent. Names that were in the index in 2010 but are no longer in 2026 are absent. Names that have been acquired, bankrupted, or otherwise removed are absent. The remaining universe is by definition the survivors.

Why this matters more for shorts

Survivorship bias works in a specific direction:

Survivors are, on average, better-performing than the original universe.
Long strategies that hold survivors show inflated returns.
Short strategies that would have shorted non-survivors miss the largest gains (since the worst-performing names often went to zero before delisting).

The asymmetric impact on short strategies is large. Studies of post-offering drift, post-merger underperformance, and other short-side anomalies show that a significant share of the realized alpha comes from names that subsequently delisted. Excluding them underestimates the strategy's true gross return.

What proper handling looks like

To handle survivorship bias correctly:

Build a point-in-time security master. For each historical date, the set of tradable securities as of that date — including names that subsequently delisted.
Keep delisted names in the dataset. Their price history continues until delisting; their final terminal value is the delisting return.
Use total returns that include delisting. For names that delisted to zero, the final return is -100% (less any recovery). For names acquired at a defined price, the final return is the deal price relative to last close.
Test strategy performance on the full universe. Including delisted names.

Delisting returns

How to value the terminal return for delisted names is a known problem with multiple defensible approaches:

Last-trade approach: Use the last reported trading price before delisting. Often understates the true terminal loss because trading frequently freezes at non-trivial prices before the actual delisting.
Post-delisting OTC approach: For names that continued trading OTC (pink sheets) after exchange delisting, use the OTC price. More accurate but requires OTC data.
Zero approach: Mark the terminal value at zero for names that delisted under distress. Conservative but defensible for bankruptcy cases.
CRSP delisting return: CRSP publishes standardized delisting returns that researchers commonly use as the canonical value.

For dilution-event research, the CRSP approach or a careful OTC continuation is typically the right choice.

Practical sources

To build a survivorship-bias-free universe:

Use a point-in-time security master that includes delisted names.
Source historical index constituents from a vendor that maintains them (not "current constituents extended back").
Use delisting-aware total return data.

Common free sources (Yahoo Finance, free APIs) typically do not provide delisted names. The cost of survivorship-bias-corrected data is meaningfully higher but is essential for credible backtests.

Estimating the impact

For a sense of magnitude: studies of the small-cap US equity universe over the last 20 years suggest that 5–15% of names delisted per year on average, with substantially higher rates in micro-cap and recently-IPO'd cohorts. The bias in 5-year and 10-year backtests of those segments is correspondingly large.

For specific event-study research like post-offering drift, the impact varies:

Studies that exclude delisted names typically report drift figures 30–60% smaller than studies that include them.
The effect is concentrated in the smallest-cap subset of the universe.
For mid- and large-cap subsets, the bias is smaller but still material.