Survivorship Bias in Historical Equity Backtests: The Cost of Terminal-Constituent Look-Ahead

Backtests frequently begin with a list of securities observed at the end of the sample and then project that list backward. This paper measures the resulting terminal-constituent look-ahead distortion. Using audited CRSP CIZ daily security data, we construct a dynamic Top-500 United States large-cap proxy running from January 2006 through December 2025. Membership for each return date is formed at the previous trading close by ranking eligible common-stock securities on market capitalization. The design contains 1,387 securities, retains every exit, and uses total returns inclusive of delisting effects. We then commit the prohibited operation deliberately: on every date we keep only those contemporaneous members that also belong to the proxy at the end of 2025. The filter leaves 221 of 500 securities at the start of 2006 and a median of 314 across the sample. It raises the equal-weighted annualized return from 10.20 to 13.66 percent, the Sharpe ratio from 0.497 to 0.653, and terminal wealth from 6.95 to 12.89 per dollar; value-weighted effects are smaller but material. A paired twenty-day circular block bootstrap places equal-weighted Sharpe inflation at 0.155, with a 95 percent interval of 0.118 to 0.198, and value-weighted inflation at 0.051. An exact identity shows that the daily equal-weighted bias equals terminal-nonmember exposure multiplied by the return spread between terminal members and nonmembers, and that its annualization of 3.06 percentage points is invariant across every exit-subgroup screen. A placebo then settles the interpretation. Running the identical experiment on a simulated market with no cross-sectional predictability, in which every security shares the same expected return and volatility and none ever fails, reproduces and overshoots each headline number. The measured distortion is therefore mechanical terminal-rank selection rather than evidence about the economics of firms that leave the large-cap universe. Results condition on the realized end-2025 cohort and on a capitalization proxy rather than on official index membership.