Survivorship bias ate my equity curve

The first time I rebuilt a US equity mean-reversion test with delisted names included, the equity curve lost its posture. The old version had a lovely habit of buying temporary losers. The new version bought some companies that were not temporarily anything. They were on their way to zero, a restructuring, a halt, a delisting, or one of those press releases where management discovers the passive voice.

That difference was not a parameter choice. It was the data admitting what happened.

Survivorship bias is the error of studying only the things that made it through the selection process. In equity backtesting, the selection process is brutal and quiet. Companies go bankrupt, merge, get acquired, drop below listing requirements, change tickers, vanish into OTC purgatory, or leave an index. A dataset that only contains today's survivors lets you buy the winners in the past while pretending the dead were never eligible.

The classic mistake is testing a strategy on today's S&P 500 constituents back to 2000. It sounds conservative because the names are large and liquid now. It is the opposite. You have selected companies that survived long enough, grew enough, or recovered enough to be in the index today.

Run a value screen on today's constituents in 2002 and you are not testing what a 2002 trader could buy. You are testing a universe pre-filtered by the next twenty years. The losers that left the index are gone. The firms that were acquired after a good run are awkward. The firms that diluted shareholders into oblivion are missing. The backtest inherits future knowledge through the membership list.

The bias is not always dramatic in large-cap momentum. It can be enormous in small-cap mean reversion, distress, low-price filters, short books, and anything that buys apparent cheapness. A stock can look cheap because the market is wrong. It can also look cheap because the accounting will be restated, the lender is about to take the keys, or the exchange is preparing a letter.

A delisting is not a missing value. It is an economic outcome.

If your data stops on the last traded day and your backtester quietly carries the last close, drops the row, or liquidates at a convenient price, you have created money. The right treatment depends on the event, the venue, and the data available, but the wrong treatment is almost always flattering. Bankrupt names do not usually give you a clean exit at yesterday's close with no slippage and no borrow issue.

This is where vendor details matter. Norgate says plainly that it specialises in survivorship-bias-free data for US, Australian and Canadian stock markets, and its data content tables spell out how delisted US securities are handled, including delisted symbols with year-month suffixes. That sort of boring metadata is exactly what you need. The point is not that one vendor is holy. The point is that a serious equity backtest needs a record of dead names, historical constituents, corporate actions, and symbol history. If the vendor cannot tell you how those work, assume the answer is bad.

Free data is useful for sketches. I use it for sketches. The danger starts when the sketch becomes evidence.

A downloaded price panel often has adjusted closes but not the adjustment history. It may handle splits but not special dividends. It may have current tickers but not old ones. It may omit names after delisting. It may backfill ETF histories in a way that makes the tradable universe appear earlier or cleaner than it was. It may not tell you whether a price is from the primary exchange, a consolidated tape, an OTC continuation, or a stale value.

You can still learn from such data. You cannot use it to size a trade unless the strategy is almost immune to those defects, and most interesting equity strategies are not. If the edge is 30 basis points a month and the data can donate 50 by forgetting failed names, the test is theatre.

There is a simple smell test. Pick a date deep in the backtest and ask what universe the strategy believed existed on that date. Then explain why each name was eligible using only data available then.

For an index strategy, that means historical index membership with entry and exit dates. For a liquidity-screened universe, it means prices and volumes for names that later died. For a fundamental strategy, it means statement data as reported at the time, not final restated values. For a short strategy, it means borrow availability if the backtest claims to short small or distressed names.

I also want to see what happens when delisted names are isolated. If nearly all the lost performance comes from including failures, the original test was not a strategy. It was a survivor selection procedure.

I trust equity backtests more when the universe is reconstructed from point-in-time membership, when delisted returns are explicit, and when symbol changes are treated as identifiers rather than string matching puzzles. I trust them more when the researcher can name the vendor's treatment of corporate actions without opening a PDF in a panic.

I trust them less when the universe is described as "the Russell 3000" with no membership source, or "all liquid US stocks" with no dead names. I trust them almost not at all when the data came from a free endpoint and the strategy buys distressed shares.

Survivorship bias is not a philosophical blemish. It is a missing row that used to contain your loss.