# Survivorship bias in backtests case study

Source: https://finzdata.com/research/survivorship-bias-backtest-case-study

Free stock datasets cover today's tickers only, omitting bankrupt or delisted companies. FinzData survivorship_audit compares results on point-in-time universe.

Research · November 2026

# Survivorship bias in backtests: a case study with real numbers

Many free stock datasets cover today's tickers only. A backtest on that list quietly drops every company that later went bankrupt, was acquired or was delisted, so the strategy never holds a loser that disappeared. This is survivorship bias. We measured how much it changes the answer, using FinzData's `survivorship_audit` tool, which runs the same strategy twice: once on the point-in-time universe (every stock that actually traded on each date, delisted names included) and once on survivors only (stocks still trading today).

## The headline

On the top 3,000 US stocks, a standard 12-1 momentum strategy returned **6.89% a year** point-in-time, but **10.55% a year** on survivors only, from January 2005 to October 2026. That is 3.66 points a year of return that never existed. Compounded, $1 grew to **$4.28** in the honest test and **$8.92** in the survivors-only test, more than twice as much. The survivors-only test left out 5,997 stocks that traded during the period.

## All four tests

| Strategy | CAGR, point-in-time | CAGR, survivors only | Bias | $1 became (PIT) | $1 became (survivors) | Sharpe PIT / survivors | Stocks dropped |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 12-1 momentum, top 3,000 US stocks | 6.89% | 10.55% | +3.66 pts | 4.28x | 8.92x | 0.37 / 0.46 | 5,997 |
| 12-1 momentum, top 1,000 US stocks | 7.18% | 8.43% | +1.25 pts | 4.54x | 5.85x | 0.38 / 0.42 | 1,741 |
| 1-week reversal, top 1,000 US stocks | 8.53% | 11.48% | +2.95 pts | 5.97x | 10.72x | 0.41 / 0.47 | 1,741 |
| Value (book-to-market), top 1,000 US stocks | 6.79% | 6.82% | +0.03 pts | 2.75x | 2.76x | 0.45 / 0.45 | 1,272 |

Each strategy holds the top 100 names by its signal, rebalances monthly and pays 10 basis points of trading costs. Momentum and reversal run January 2005 to October 2026; the value test covers a shorter window of about 15 years, set by when book-to-market data is available in this test.

## What the numbers show

- **The bias grows with the universe.** Momentum on the top 1,000 stocks was inflated by 1.25 points a year; on the top 3,000 by 3.66. Smaller companies delist more often, so a broader universe drops more failures.

- **Strategies that buy losers are hit hard.** One-week reversal buys recent losers, and some of those later delisted. Survivors-only data removed them, adding 2.95 points a year: 8.53% became 11.48%.

- **Some strategies barely move.** The value test changed by only 0.03 points a year. The bias depends on whether the strategy tends to hold names that later disappear, so it has to be measured, not assumed.

- **The benchmark is biased too.** The equal-weight top-1,000 benchmark returned 7.36% a year point-in-time and 8.67% on survivors only. A survivors-only test can flatter both the strategy and the yardstick it is compared with.

- **Momentum on its own did not beat the market here.** Point-in-time, 12-1 momentum on the top 3,000 returned 6.89% a year against 12.85% for the equal-weight benchmark, a shortfall of 5.96 points. On survivors-only data the shortfall shrinks to 3.80 points, which is how a weak idea can look better than it is.

## Run it yourself

Every number above came from one call each to `survivorship_audit` on FinzData snapshot v2026-11-05. Ask Claude, ChatGPT or Cursor with FinzData connected: *"Run a survivorship audit on 12-1 momentum, top 3,000 US stocks, 2005 to today."* Run IDs for replay: d48c047ede234986, 740e8ca996d082a5, d9fea742006ca903, d75ad94b883e327c.

FinzData covers 16,000+ US stocks that have traded on NYSE and Nasdaq since 2000, delisted included, with every data tool pinned to an as_of date. [See the full history](https://finzdata.com/shop) or [connect FinzData to your AI](https://finzdata.com/mcp).

Historical simulations, not investment advice. Past results do not predict future returns.