Survivorship, and the companies that are not in your list

Course 2 said a screener cannot ask what would have happened, because it works on today's constituents. This lesson puts a number on that, and the number is larger than almost anyone expects.

Count the dead

Two requests to the same endpoint today, one with delisted=1:

The delisted outnumber the living by roughly 1.8 to 1.

A backtest run on "the US stocks I can list today" silently excludes 32,854 companies. Not a rounding error, not a tail — the majority of everything that ever traded. And the excluded set is not random: it is disproportionately the failures, the bankruptcies, and the ones that fell far enough to be delisted.

Why this makes bad rules look good

Every company in your universe is one that made it to today. So the backtest quietly assumes that whatever you bought, it survived. "Buy the biggest drawdowns and hold" looks superb on survivors, because survivors are by construction the names whose drawdowns did not end in a delisting. Not that every one recovered — Cisco closed at 144.38 in March 2000 and was still near 109 in September 2026, twenty-six years later on price — but none of them went to zero. Run it on the full set and the trades that went to zero come back, and they were always the majority of that strategy's exposure.

The bias is strongest exactly where beginners look: deep value, mean reversion, buying weakness. Those strategies are, structurally, bets on recovery — and a survivor universe has pre-selected for recovery.

What you can actually do about it

The full fix — a point-in-time universe, reconstructing membership as it was on each date — is real work and beyond this course. Three things that are not:

Include the delisted set. The list is one request away. Your backtest can at least run over live-plus-delisted rather than live-only, and comparing the two results is the honest experiment.

Report the universe as an assumption. "Ran on 18,012 live US common stocks as of 2026-08-25; excludes 32,854 delisted" belongs in the report, beside the parameters. A reader can then judge the result rather than trust it.

Prefer single-instrument or index-level tests while learning. A backtest on SPY has no survivorship problem worth the name; a cross-sectional stock-picking backtest is riddled with it. Know which one you are running.

The same bias in a different costume

Two relatives worth naming, because they come from the same habit of using today's list:

• Index reconstitution. Backtesting "the S&P 500" using today's members means holding companies that were added after they had already performed well. The index is rebalanced towards winners, and using today's membership backdates that.
• Ticker reuse. A symbol freed by a delisting gets assigned to a different company later. Stitch the price history by ticker and you have concatenated two unrelated businesses into one series.

The finance behind it

Where the delisted tickers live as data: Where do you find the tickers that no longer exist?

Try it now

The delisted list is the course 2 universe request with one parameter more: /exchange-symbol-list/US?delisted=1&type=common_stock through your proxy. Its first rows today:

Count the whole list and set your number beside the table above; that table is dated 2026-08-25, and on 28 September 2026 the same request returned 33,112 rows. Then pick five delisted tickers and read their last year of prices. Here is what one death certificate looks like, SVB Financial Group, first under its NASDAQ code:

and then under the code it traded as afterwards, with the same ISIN:

A backtest that knows only SIVB marks the position at 106.04 and stops. The first session with any volume under SIVBQ, 28 March 2023, closed at 0.40. Look for the same two things in your five: where the series ends, and whether the company went on trading under another code. Reading five real death certificates does more for your judgement than any explanation of the bias, and it takes ten minutes.