Historical S&P 500 constituents and survivorship bias

How it works

An index is three things stacked on top of each other: a rulebook, a brand, and a membership history.

The rulebook is published — domicile, listing venue, float, liquidity, profitability — and you can read all of it and still not say who was in the index in March 2011. The brand is a trademark, licensed to fund issuers and exchanges. The membership history is the third thing, and it is sold rather than published: which securities were members on which dates, plus the corporate-action-adjusted return series computed from them.

That is why survivorship bias is a pricing problem rather than a statistics problem. Everybody knows the concept; what is rarely said is that the fix is a purchase, because the only complete record of who was in the index and when is kept by the company that maintains it.

A backtest run on today's constituent list is not a broken version of the right experiment. It is a clean test of a different hypothesis: stocks successful enough to still be in the index in 2026 performed well between 2000 and 2026. That is guaranteed true, and a strategy tested against it will look excellent for reasons that have nothing to do with the strategy.

Point-in-time means that on every date in the test you hold the securities a rule-following investor could actually have held that day, using only what was knowable then. Two clocks matter and they are not the same clock. Coinbase Global was announced as a member of the S&P 500 on 12 May 2025, replacing Discover Financial Services because Capital One was acquiring it, and the change took effect before the open on 19 May 2025. A reconstruction that stamps membership at the announcement date carries a week of look-ahead, in a name that moved hard on the announcement.

Reconstructing this from press releases is harder than it sounds, because one corporate action is rarely one membership change. Honeywell's aerospace spin-off announcement of 23 June 2026 moved Honeywell Aerospace into the S&P 500 and the S&P 100, pushed Conagra Brands out of the 500 and down into the S&P SmallCap 600, removed two names from the 600 to make room, and renamed the parent to Honeywell Technologies while leaving it in the 500 — effective across 29 June, 30 June and 1 July. Company identity has to survive a rename, index identity has to survive a demotion, and every leg needs its own effective date.

Share classes are the other quiet failure. Index members are securities, not companies. The SPDR S&P 500 ETF Trust's daily holdings file for 11 September 2026 carries Alphabet twice — class A under GOOGL, CUSIP 02079K305, and class C under GOOG, CUSIP 02079K107 — with separate weights. A membership table keyed on company gets the weights wrong. A table keyed on ticker breaks the first time a ticker moves.

The dead are the hard part

The survivors are easy. Every free data source has Apple's price history back to 1980. What almost none of them have is the company acquired at a discount in 2008, the one that filed Chapter 11 in 2015, or the one taken private and gone — the half of the universe containing the losses. A backtest that cannot buy a stock that went to zero cannot lose money on it.

Two things break the moment a company delists. The price history often disappears from cheap feeds entirely rather than ending, so the absence looks like a coverage gap rather than an event. And the symbol is recycled: an exchange can reassign a three-letter ticker to an unrelated company years later, so any table keyed on ticker silently splices two businesses into one price series. That is what a permanent security identifier is for, and it is the most useful question to put to a vendor.

When a vendor claims survivorship-bias-free, ask precisely this: does a delisted security keep its permanent identifier and its full price history inside the same tables as everything else, up to the last print — or does it only appear in a separate delisted file, with the history cut short at the delisting? The two claims sound identical in marketing copy and produce very different backtests.

Why the membership data is worth money

Index licensing is priced off assets, not off effort, and passive flows have made that an unusually profitable arrangement. The numbers below come from the providers' own filings.

S&P Global reported its Indices segment for 2025 at $1,850 million of segment revenue — $1,839 million of it from external customers, the rest intersegment — and $1,271 million of operating profit, a 69 per cent operating margin and 71 per cent on the company's adjusted basis, in its Form 10-K filed on 11 February 2026. Of that revenue, $1,206 million, or 65 per cent, was asset-linked fees charged on money held in ETFs and mutual funds tracking the benchmarks; data and custom index subscriptions were $320 million, 17 per cent.

MSCI's Index segment reported $1,786.8 million of revenue for 2025 and adjusted EBITDA of $1,366.0 million — a 76.4 per cent margin, on total company revenues of $3,134.5 million — in its fourth-quarter release of 28 January 2026. LSEG's FTSE Russell division reported £954 million of revenue and £635 million of adjusted EBITDA, a 66.6 per cent margin, in its 2025 preliminary results. Nasdaq's Form 10-K counts 451 exchange-traded products tracking its indexes, holding $882 billion at 31 December 2025.

The mechanism, without any claim about intent: the fee scales with assets in tracking products, while the cost of maintaining the rulebook does not. The same customers then buy the membership data as a subscription, and the historical series is inventory already owned.

The ownership map, briefly. S&P Dow Jones Indices LLC is consolidated by S&P Global, with minority partners CME Group and CME Group Index Services holding 27 per cent and a 2012 licence agreement giving the joint venture a share of profits from CME's equity index products — $193 million of revenue in 2025. MSCI Inc. is standalone and listed. FTSE Russell is one of LSEG's four divisions. Nasdaq runs its index business inside Capital Access Platforms. CRSP, the academic standard built at the University of Chicago, was acquired by Morningstar for $365 million, completed on 2 February 2026.

None of the five sells point-in-time constituent history to an individual developer at a published price; CRSP is licensed to academic institutions, government agencies and investment practitioners rather than sold at retail. What an indie buys is a redistributor's repackaging, and the price band below is that market, not the index providers'. Where index licensing sits in the industry as a whole is mapped in who owns the market data industry.

What it costs

Norgate Data is the cheapest honest number: US stocks at $630 a year on the Platinum tier, where delisted securities and index constituent membership for the S&P, Russell, Nasdaq and Dow families begin. The $270 Silver and $360 Gold tiers are currently-listed securities only, so the feature the product is bought for is absent from both. It is Windows-only and rented — the database locks when the subscription lapses.

AlgoSeek carries point-in-time index membership alongside tick-level history, leased per dataset at published monthly prices — the option when the test needs intraday detail rather than daily bars.

FirstRate Data sells once rather than renting: 16,304 US tickers from January 2000, roughly 7,000 of them delisted and suffixed so a reused symbol cannot collide. The vendor states that this is not the complete universe and that coverage thins with age — the right disclosure, and also why it is cheaper.

Kibot answers the survivorship question in its own FAQ with "partially", driven by how liquid a name was when it stopped trading. Believe the word. QuantRocket bundles US history back to 2007 with delisted names retained into a self-hosted Docker stack, if the data and the engine are being bought together.

Free and cheap routes, and where each one breaks

Wikipedia page history is the first thing everybody finds and the weakest thing here. It is an editing log, not a record: it says when somebody edited a page, not when a change took effect. Edits arrive late, arrive early, get batched, get reverted. Use it to discover that a change happened, then date it from the provider.

The index provider's own announcements are authoritative and free, and each carries both dates — exactly the pair a point-in-time table needs. What they are not is a table: one press release per event, reaching back no further than the archive does.

SEC filings tell you a merger closed. They do not tell you what the index did about it, or when.

ETF holdings files are the genuinely underused route. Issuers of the large S&P 500 trackers publish complete daily holdings as a dated file — the SPY file cited above is one spreadsheet at one URL. It is the fund's holdings rather than the index, so expect a cash line and small timing differences, but for membership it is close enough for almost any backtest. The catch is direction: archiving works forwards only. The rest of what is free because a regulator publishes it is in free stock market data.

What you can do about it

Pick the branch you are actually on.

Academic work, or anything that will be reviewed. Use your institution's CRSP subscription for the security master and the delisted history, and its index-provider licence for membership. You are not buying this personally, and a reconstruction will not survive a referee.

Testing US large caps, and able to accept a known bias. Keep the current constituent list, then write the sentence: this test uses index membership as of the run date and is therefore biased upward by an unmeasured amount. Put it in the results, next to the Sharpe ratio. A stated bias is a limitation; an unstated one is a claim.

You need real point-in-time membership. Budget $630 a year for daily bars with constituent history, more for intraday. Before paying anyone, ask the identifier question above and check that constituent history is in the tier you are being quoted, because on at least one vendor it is not. The rest of the field is in market data APIs.

You are building something that runs for years. Start archiving the daily holdings file today: a cron job, a dated filename, a cheap bucket. In two years you own a point-in-time membership series nobody licensed to you. Do it alongside whichever branch above applies, not instead of it.

Free scraped lists, of the kind yfinance users assemble, are fine for a study project and not for anything with money behind it. The engines are agnostic: every tool in backtesting frameworks will run a biased test at full speed and report a beautiful number for it.

Whatever you pick, re-run the old backtest on the new data before trusting either. The correct outcome is that your best strategies get noticeably worse. If nothing changes, the new data is not doing what you paid for — better found out in the first week than in the first drawdown.