Guide · Research method · Updated 2026-09-23

A list chosen today is a list of winners, and on the exchange's own files that choice added 3.6 to 10.5 points to a 2022 to 2026 backtest

Survivorship bias already has its entry on this site's own checklists, in the standard list of backtesting mistakes and the eight ways a backtest lies. Naming it is the easy half. Its size depends on the universe, on the window and on what a company that disappeared was actually worth to the people holding it, and all three can be measured from files the exchange publishes after every session.

What follows is that measurement, with every rule stated so it can be rerun. The findings that matter most are the ones that complicate the textbook version: a symbol is not a security, a security can leave one exchange's file without leaving the market, and the most common way of choosing a test universe is a far larger error than delistings.

Selection on the outcome, written as an identity

A list compiled at the end of a test period is conditioned on having lasted to the end of it. A backtest on that list therefore reports the average return of names given that they survived, which is a different quantity from the average return of the names a trader could have held at the start. The gap between the two has an exact form. If a share of the starting list is missing from today's list, the survivors-only result exceeds the full result by that share multiplied by the difference between what the survivors became and what the missing names were worth. Nothing in that product is an opinion, and neither factor is zero in a market where companies merge, delist and get suspended.

On the list measured here, 52 of the 1,119 names, 4.6 per cent, print no row on 18 September 2026. The survivors turned ₹1 into ₹2.30 on average. The missing names were worth ₹1.08 at their last traded price, or ₹0.04 if every name that stopped trading is valued at zero. The product is 5.7 points in the first case and 10.5 in the second, and the build asserts that it equals the difference between the two portfolios to nine decimal places.

The second factor is large even when an exit loses nothing, because an exit stops compounding. The median name that stopped trading left at 1.07 times its starting price; the median survivor ended at 1.47 times and the mean survivor at 2.30. In a window in which the survivors compounded at 19.4 per cent a year, a name frozen at a fair price still falls behind. The same exit share costs more in a strong market than a weak one, which is one reason the estimate moves with the window.

The list you could have traded on 3 January 2022

The first session in the files is 3 January 2022. The full security bhavcopy lists 1,660 rows in the EQ series that day, and EQ is not a list of companies: 115 of them are units of exchange traded funds. The full bhavcopy carries no ISIN, so the cash bhavcopy for the same session, which does, supplies it, and every row whose ISIN begins with the fund prefix is removed. That leaves 1,545 company shares. The stated liquidity filter keeps those that traded at least ₹1 crore that session, 1,119 names, using nothing but that session's file. The notional portfolio buys each at that close in equal rupee amounts.

Each name is then followed to 18 September 2026 across symbol changes, and its price is adjusted for splits and consolidations by the face value ratio, bonuses by the share ratio, rights issues at their theoretical ex-rights price when the issue was priced below the market, and dividends reinvested on the ex-date. Demergers and similar distributions, 46 of them on this list, are taken at the parent's opening price on the ex-date against its last close before it, because the file's previous close is not adjusted on those days; the guide to adjusting a price series sets out why. Two overnight moves outside a band of 0.70 to 1.40 had no listed event and were treated as unlisted capital changes. The 246 company shares that were in the trade-for-trade series BE and BZ that day are measured separately, because that is where the bias concentrates.

A symbol is a label, and 70 of the exits were labels changing

Of the 1,119 names, 997 print a row on the final session under the symbol they had on the first. A test keyed on symbols stops there and books the other 122 as exits. Of the whole list, 71 names changed symbol, 66 of them on the exchange's symbol change list and 5 found only by the previous-close bridge, which pairs a new symbol whose first previous close equals the last close of a symbol that stopped within five sessions; every bridge was checked against the issuer code inside the ISIN. All but one of the renamed companies still trade. Nine more names printed no row on the final session without having stopped: 5 traded within the final month, 3 of them only in the weekly window allowed to a suspended company, and 4 had been withdrawn from this exchange while listed on another. That leaves 43 genuine exits, and a symbol-keyed count overstates them 2.8 times, the same effect the bhavcopy field guide measured across every symbol in the cache.

The error is not neutral. The 70 renamed survivors turned ₹1 into ₹3.31 against ₹2.23 for survivors that kept their symbol, so a symbol-keyed survivors-only result is too low, at ₹2.23, and its full-list version with exits at zero is far too low, at ₹1.99. Keyed on symbols, the bias reads 5.6 to 24.3 points, and the harsh end is 2.3 times the true one.

The ISIN is not a permanent key either. A rename usually keeps it, as 63 of the 71 renames here did to the end of the window, but a new face value brings a new ISIN: 153 companies that never changed symbol carried a different ISIN at the end, and every one of them had a split or a consolidation. A key that survives both events combines the issuer code in the ISIN with the corporate actions list.

The measured bias, and the bracket around it

The daily file records that a name stopped trading. It never records why, and the routes pay very different amounts. A merger usually pays in the acquirer's shares at the scheme's swap ratio. A voluntary delisting under the SEBI (Delisting of Equity Shares) Regulations, 2021 pays the price discovered by reverse book building or, for announcements from 25 September 2024 in frequently traded shares, a fixed price at least 15 per cent above the floor price. A compulsory delisting obliges the promoters to buy the public's shares at a fair value set by an independent valuer appointed by the exchange. A suspended company can sit in the file's weekly window for months and then stop printing. None of these terms is in the file, so the full-list result is reported under each stated exit value, from the most generous to the harshest.

The generous end values each exit at its last traded price and reinvests the proceeds across whatever was still held on that session, never across the eventual survivors, which would itself look ahead: 3.6 points. Holding the proceeds in cash instead gives 5.7. Valuing every exit at zero gives 10.5. One observable narrows the range. The 35 exits last seen in the EQ series left at a median 1.20 times their starting price, the profile of a merger or a delisting at a market price; the 8 last seen in BE or BZ left at a median 0.20 times. Valuing the first group at its last price and the second at zero gives 3.8 points. The window is 4.7 years, so the whole range is 0.39 to 1.18 points a year, on a list that compounded at 19.4 per cent a year before the correction.

For scale, the exchange's own price indices rose 1.50 times for the Nifty 500, 1.83 times for the Nifty Smallcap 250 and 2.52 times for the Nifty Microcap 250 over the same window, before dividends. An equal-weighted list of 1,119 names, most of them small, belongs between the last two.

Less traded, more to hide

Split all 1,545 company shares in EQ on the first session into fifths by that session's traded value, and run the same arithmetic inside each fifth.

Both factors of the identity grow as trading thins. In the least traded fifth, 28 of 309 names, 9.1 per cent, are missing from the final list, against 11 of 309, 3.6 per cent, in the most traded fifth, and the bias runs 5.5 to 23.7 points against 0.9 to 7.0. The middle three fifths sit within a few points of one another, so the gradient lives at the ends rather than climbing evenly. The names that began in the trade-for-trade series are a different population: 67 of 246, 27.2 per cent, are missing at the end, the missing ones were worth ₹1.11 at their last price, and the bias is 25.0 to 57.1 points. A backtest of thinly traded or low-priced shares built from today's list starts with an error of this order.

One pattern in the table deserves suspicion rather than celebration. The survivors of the least traded fifth averaged ₹2.71 against ₹2.02 in the most traded, yet their median was lower, 1.37 against 1.51. The mean of a thin list is carried by a few names that multiplied many times, and an equal-weighted book earns the mean. That is exactly the kind of result the multiple testing problem feeds on.

The estimate moves with the window

The bias accumulates with time in two ways at once: more names leave, and the survivors have longer to compound away from them. From 2 January 2023 the range is 1.2 to 6.7 points; from 1 January 2024 it is 0.5 to 2.6; from 1 January 2025, with 11 exits, it is minus 0.04 to 0.9, and at that length the generous end is noise. A one-year backtest shows almost none of the problem, and a ten-year backtest on the same universe would show more than any row here. The fund units removed rise from 115 to 226 across the four start dates, so a filter on the EQ series alone admits more non-companies every year.

The exchange's file is not the market

Rows can stop and start without any listing ending or beginning. The exchange's permitted-to-trade category lets shares listed on another exchange trade on this one. After a review announced in circular NSE/CML/2023/24 of 29 March 2023, the exchange withdrew 11 such securities with effect from 2 May 2023 (NSE/CML/56237), extended 3 of them whose companies had applied to list (NSE/CML/56527, which let them trade until their due diligence was complete or for 90 days, whichever came first), and withdrew 10 more with effect from 26 October 2023 (NSE/CML/58560). In the files, 16 of the 18 withdrawn names printed their last row within three sessions of the effective date. Their companies were listed on another exchange, which is what the category means, so none of those last rows was a delisting; a test keyed on this file alone books every one of them as an exit.

Then the route reopened. Circular NSE/CML/73797 of 17 April 2026 admitted 106 securities of companies permitted to trade to dealings from 20 April 2026, under Regulation 3.1.1 of the exchange's capital market trading regulations. On that session 92 symbols printed their first row anywhere in the files, 89 of them from the circular, and the exchange's list of listed equities gives 103 companies a date of listing of 20 April 2026, which is the date they joined this file rather than the date they listed. Of the 18 names withdrawn in 2023, 12 are back, 6 of them on that same day. A method written before April 2026 that reads a symbol's first row as a new listing is wrong for every one of them, and a universe pulled from today's file contains 106 companies whose history on this exchange has gaps of years.

Suspension has its own signature. SEBI's master circular of 11 November 2024, as applied in the exchange's circular NSE/CML/65586, permits a suspended company's shares to trade on a trade-for-trade basis in series BZ on the first trading day of every week for six months, after which that window closes. The file shows it plainly: over the final quarter, BZ printed 37.8 rows on the first session of a week against 27.1 on other sessions, and 9 symbols printed only on those sessions. Three names on this list sit in that window, suspended but still priced once a week, which is why names still printing in the final month are valued at their last price rather than counted as exits. The trade-for-trade series themselves are covered in the surveillance frameworks guide.

Survivorship's cousins: the index list and the fund list

An index's published daily levels are point in time, at least from its launch, because each day's level was computed from that day's members. A list of its current members is not. The Nifty 500 is reviewed twice a year on six-month average data to 31 January and 31 July, with four weeks' notice, and a member is excluded when its rank by full market capitalisation or by average turnover falls below 800, according to its factsheet of 31 August 2026. A backtest on today's members inherits every one of those decisions taken after the test began.

The size of that cousin can be measured with a proxy. Rank the same 1,545 companies by traded value twice, once as a trader could on 3 January 2022 and once as a reader does today, by mean traded value over the final 21 sessions, and hold the top 500 of each in equal amounts. The list available at the start turned ₹1 into ₹2.06 to ₹2.00, with a delisting bias of its own of 3.1 to 9.4 points. Today's list turned it into ₹3.58. Only 322 names are on both. Choosing the list today adds 1.52 to 1.58 per rupee, more than 16 times the delisting bias on the same list, because traded value rises with price and ranking on today's value selects on the outcome almost directly. An index ranked on today's market capitalisation does the same by construction.

Fund categories repeat the pattern. SEBI's circular SEBI/HO/IMD/DF3/CIR/P/2017/114 of 6 October 2017 allowed one scheme per category, with exceptions for index funds and exchange traded funds tracking different indices, fund of funds with different underlying schemes and sectoral or thematic funds, and fund houses proposed winding up, merger or a change of fundamental attributes for their similar schemes. A category average computed today over surviving schemes leaves out the records merged away then. SEBI's circular of 26 February 2026 revising categorisation discontinued the solution-oriented category, stopping subscriptions and requiring those schemes to merge into schemes of similar asset allocation, and gives sectoral and thematic schemes three years to bring their overlap with other equity schemes, large cap schemes apart, to 50 per cent or less or be merged. Category histories drawn from a scheme list pulled after those mergers lose those records too. The same identity governs the error; no fund data is measured here. The same filter applied to strategies, accounts and publishers is measured in reading a published strategy claim.

What a point-in-time universe requires

A point-in-time universe is a list of securities that is itself a dated fact: on every date it holds what could have been traded on that date and nothing else. For a buy and hold from one date that is the starting list. For a strategy that selects from its universe each period, it is a fresh list on every rebalance date, drawn from that date's file, which is where a current-members shortcut does its damage.

Two details decide whether the construction is right, and most descriptions leave both out. The corporate actions list files past events under the symbol a company uses today, so a demerger from 2022 appears under a symbol adopted in 2024, and collecting events by the symbol a name had at the start silently drops them; they have to be gathered under every symbol a security used. And an ISIN that changes across a long suspension with no split on the list is the mark of a restructured share capital, typical of an insolvency resolution, where the relisted price is not the old holder's price. Five names in the full starting set carried that mark and one more a capital reduction on terms the list does not give; each is valued as an exit at its last price before the event. The whole construction, with its inputs and rules, is written down so it can be rerun, which is the standard set out in publishing a checkable result.

What survives the correction

The level survives. After correction the liquid list compounded at 19.0 to 18.2 per cent a year against 19.4 before it, so the statement that an equal-weighted book of these shares more than doubled between 3 January 2022 and 18 September 2026 holds under every exit value tested, gross of costs and taxes.

The ranking by traded value survives, narrowed. The least traded fifth still led the most traded fifth under every assumption, ₹2.65 to ₹2.47 against ₹2.01 to ₹1.95, but the lead shrinks from 0.69 per rupee to between 0.52 and 0.64, and it belongs to the mean. A claim about the typical thinly traded stock was never supported: the median survivor of that fifth, at 1.37, trailed the most traded fifth's 1.51 before any correction.

Three results do not survive at all: anything measured on names that began in the trade-for-trade series, where the correction is 25.0 to 57.1 points; an edge of a point or two a year measured on a thin universe, because the bias in the least traded fifth is 0.5 to 2.4 points a year on its own; and anything measured on today's most traded names, where the selection alone is worth 1.52 to 1.58 per rupee. Survivorship is one of three ways a backtest can beat its benchmark by construction; an unfair comparison and a search across many variants are the other two, taken apart in choosing the right null and the guide to multiple testing. Treating a universe as a dated fact rather than a download is method, and it is taught as method.