Data to 20 November 2026
About

Methods

ETF desk: method

Code: pipeline/tipsheet/compute/etf.py (build) and publish/etf.py (bundles). Shared AMFI readers: compute/amfi.py. Tests: pipeline/tests/test_etf.py.

What is measured

For every exchange-traded fund listed on NSE, day by day, as far back as both the exchange price and the AMFI NAV exist:

Sources

WhatDataset (dataset)From
Closes, traded value, ISINnse_bhavcopy and its eras, series EQ2002 for the oldest ETFs; ISINs printed from 2011-06-22
NAVamfi_nav_history and its eras, amfi_nav_latest2006-04-03
Benchmarksnifty_index_total_returns via compute/indices.load_tri (bad prints repaired)index launch dates
Gold benchmark (rough)domestic gold price in rupees from asset_tri_daily.parquet2005
Distributions and splitsstocks_events_corporate_actions (ETF ISINs)2001
TERamfi_ter_daily*2018-04
AUMamfi_aum_schemewise (quarterly average, rupees lakh)2006
AMFI’s own tracking differenceamfi_tracking_difference2022-05
Order book (best bid and offer, five levels), intradaykite_etf_quotes (Kite Connect, laptop, internal only)2026-10-09

How ETFs are matched to AMFI schemes

  1. The universe. Every NSE security in series EQ with a mutual-fund ISIN (INF…). Before NSE printed ISINs, a symbol’s earlier rows take the symbol’s first ISIN, if that ISIN was already trading in the first month ISINs were printed.
  2. By ISIN. The ISIN appears in AMFI NAV history or NAVAll.
  3. By symbol chain. The same NSE symbol continues under a later ISIN that did match. ISINs change at unit splits and AMC mergers.
  4. By NAV fit. The close tracks one AMFI ETF scheme at a constant unit factor on at least 90% of 20 or more overlapping days, within ±3%.
    • Candidates from the same AMC are tried first. The AMC is identified from the ISIN’s issuer prefix.
    • If two schemes fit about equally well (for example, gold ETFs from different AMCs priced per gram), the ISIN is left unmatched rather than guessed.
  5. One ETF can span several AMFI codes. ISINs that share an AMFI code, or follow each other on one symbol, form one ETF. Its NAV is assembled from all those codes in time order.
    • AMCs re-code schemes when they merge. The BeES funds went from Benchmark to Goldman Sachs to Reliance to Nippon, and each step brought a new AMFI code.
    • Remaining traded days without a NAV are filled the same way, by NAV fit against AMFI ETF codes no other ETF has claimed. The 2006–2011 Benchmark codes carry no ISIN at all.

Result (2026-09-30):

Unit splits

ETFs split their units, and Nippon’s in particular have. AMFI and the exchange do not always apply a split on the same day.

Tracking difference

Agreement with AMFI’s own disclosure.

When equivalent AMFI scheme-name spellings exist, the spelling with the most recent reporting period is used; equal-period ties use a stable name order. This prevents a historical spelling from replacing a current disclosure merely because database rows arrive in a different order. The figures below are the recorded validation sample, rather than a live count of every subsequent refresh.

Caveats

Bundles

Logical nameWhat it is
etf/universeOne row per live ETF: category, benchmark, AUM, TER, 1- and 3-year tracking difference (mine and AMFI’s), median premium over the last 60 sessions, latest premium, median traded value (20 and 60 sessions), first date with both price and NAV, whether a detail page exists. Compact table (columns + rows)
etf/premium/<group>_dailyequity, gold, silver, debt, liquid, international. Each session: the cross-sectional median premium across the group’s ETFs that traded with a same-day NAV, with quartiles and the number of ETFs (weekly until 2026-10-09)
etf/detail/<symbol>Top 40 non-liquid ETFs by traded value. Daily: premium at the close and at the day’s average price (traded sessions only), traded value and its 20-session median, 1-year tracking difference at month-ends. meta.splits lists detected splits (weekly until 2026-10-09)
etf/liquidity/by_groupTraded value a session by kind (equity, gold, silver, other debt, liquid, international), 20-session average, from 2011-06-22
etf/liquidity/concentrationMonthly: the share of ETF traded value in the most traded one, five and ten ETFs; ETFs that traded; the share that traded on fewer than half the month’s sessions
etf/density/same_indexIndices tracked by three or more live domestic ETFs (gold and silver by metal): the most traded ETF’s share of trading and of assets and its TER, against the cheapest ETF’s
etf/density/liquidity_premiumEquity ETFs: median traded value against the 5th-to-95th percentile range of the daily premium over 250 sessions
etf/intraday/spreadsOnce five sessions are captured: per ETF, median quoted spread, two-sided share and five-level depth
etf/stats/premium_median_{equity,gold,silver,international}Stat bundles of the daily median premium
etf/stats/traded_value_all_etfs_20dTotal ETF traded value, 20-session average, Rs crore a session
etf/stats/td_1y_median_equityMonthly median 1-year tracking difference across equity ETFs

Derived tables in .cache/derived:

What the data says (2026-09-30)

What holders got: price against the index (added 2026-10-02)

AMFI’s tracking error for index funds and ETFs (etf/tracking_error/*, added 2026-10-08)

AMFI publishes each index fund’s and ETF’s tracking error every day: the annualised spread of the daily gap between the fund’s return and its index’s, over the past year. SEBI caps it at 2%. These charts use AMFI’s figures as published (Data bank amfi_tracking_error, from June 2022), unlike the ETF measures above, which are computed here from prices and NAVs. Code: pipeline/tipsheet/compute/tracking_error.py; tests: pipeline/tests/test_tracking_error.py.

How the figures are prepared:

What it shows (30 September 2026, 644 schemes):

Daily cadence, Kite and NSE’s ETF page (2026-10-09)

What was weekly. The data never was. AMFI NAVs (amfi_nav_history, amfi_nav_latest) are fetched every evening and the bhavcopy every session; etf_daily.parquet holds one row per ETF per session with a same-day NAV on 98.5% of traded days. The published bundles were downsampled: etf/premium/*_weekly took the median of each week’s daily medians, and etf/detail/* took weekly medians, so the page’s charts were weekly. They are daily now. (etf_issuer_nav_history is unrelated: US ETF issuers’ NAVs for the global page.)

Kite’s daily ETF prices add nothing to the bhavcopy (audit audits/kite/a8_etf_daily.py, 2024-10-09 to 2026-10-08, 129,571 ETF-days):

What Kite does add is the live order book, which no public source keeps. kite_etf_quotes captures the best bid and offer and the five visible levels for every ETF at 10:00, 12:00, 14:00 and 15:20 IST on sessions (laptop timer tipsheet-etf-intraday, scripts/etf_intraday.sh). Only derived figures are published: quoted spread in basis points, the share of snapshots with both sides quoted, and five-level depth in rupees lakh.

NSE’s ETF page (/api/etf) has no history to scrape: it is a live table that freezes on a 16:00 copy after the close. The historical record of ETF trading is the bhavcopy, already used here; it agrees with the page’s 8 October total of Rs 3,666 crore to 0.004%.

Intraday measures (compute/etf_intraday.py, published once five sessions are captured):

Audit findings (2026-10-09)

  1. Untraded sessions were not counted as zero, though this note said they were. The bhavcopy has no row for an ETF that did not trade, so the 20- and 60-session medians spanned more than 20 or 60 sessions. Fixed by laying each ETF on the session calendar. Today it changes 6 live ETFs’ medians (the 60-session median falls by 1.1 to 9 times); historically, when many ETFs traded rarely, much more.
  2. Three wrong “AMFI’s figure” matches and one wrong TER. The name matcher’s fuzzy step crossed funds: HDFC Nifty Bank ETF took the HDFC Nifty PSU Bank ETF’s tracking difference (-0.48), and SBI’s Midcap 150 and Smallcap 250 ETFs were matched to DSP’s. HDFC Nifty Private Bank ETF took DSP Nifty Private Bank ETF’s TER, because the HDFC name appears in two spellings in AMFI’s file. A close spelling is now accepted only when the names differ in spacing and filler words; duplicate spellings take the latest.
  3. Checked and correct:
    • NAV dates: every equity, gold and silver ETF’s close fits the same day’s NAV better than the day before or after (2024-10 to 2026-10).
    • Bhavcopy volumes and values against NSE’s ETF page: equal volumes for all 351 traded ETFs on 8 October.
    • Benchmark names against AMFI’s own benchmark field: the 4 mismatches out of 203 are spellings (”&” against “and”).
  4. Bharat Bond and gilt ETFs fit the previous day’s NAV marginally better than the same day’s (standard deviation 0.15 against 0.16 points). The difference is within noise; it reflects bond valuation timing, not misdated NAVs. Left as is.

Liquidity findings (2026-10-08)

ETFs on BSE and other non-NSE indices: checked, not built (2026-10-09)

49 of the 261 live equity ETFs have no tracking difference here. They track BSE indices (Sensex, BSE 500, Bharat 22, PSU Bank, Power, India Defence…), MSCI India, or NSE indices missing from the total-return data. Two ways to cover them were tested and rejected.

Turnover periods and concentration (clarified 16 October 2026)

Monthly concentration ranks ETFs afresh within each calendar month by the sum of NSE traded value. The top-one share is that month’s largest ETF total divided by all ETF turnover in that same month, including liquid and international ETFs. The leader can change between months. The latest point is month to date through its printed session date; it is not a completed-month forecast or an arithmetic mean of daily shares. Historical leader identities are retained in etf/liquidity/concentration metadata; the numeric series also carry the leader’s turnover, total turnover and session count.

etf/liquidity/latest publishes every ETF observed in the latest month with its latest-session turnover and share, month-to-date turnover and share, and average turnover per observed market session. Funds that stopped trading during the month remain in the denominator. A missing latest-session row counts as zero turnover; a zero-turnover session has no defined share. The metadata prints the first/last dates, number of sessions, and both denominators. This reconciles the monthly concentration plot with a same-period ranked comparison.

The category turnover chart and the total-turnover reading use a 20-session mean. Per-fund comparison tables use 20- or 60-session medians, labelled by window. These medians cannot be added to obtain market turnover or divided by the market mean to recover a monthly share. Turnover is exchange activity, not subscriptions, redemptions or net money entering ETFs.

Read-only source reconciliation: audits/etf/check_turnover.py. It uses the local derived cutoff, independently reads current-month NSE source rows, checks every turnover cell, and reconciles total shares and top-one/five/ten shares.

ETF comparison and individual histories (16 October 2026)

compute/etf_insights.py and publish/etf_insights.py add etf/insights/comparison and one on-demand profile per active ETF, keyed by its stable entity ISIN. The density step refreshes these with the existing ETF inputs. Profiles contain only premiums, return gaps, fees and asset statistics, not licensed raw price or index levels.

Same benchmark. Domestic fund benchmarks use the computed match, falling back to AMFI’s disclosed benchmark. Spelling normalisation follows the existing tracking-error matcher; gold and silver are grouped by metal. Funds without a matched benchmark remain browsable but are not ranked as peers. Missing return comparisons stay absent; the page prints the number available for each metric. There is no aggregate score or implied recommendation. Gold’s computed return gap retains its approximate-benchmark caveat; overseas premiums retain the foreign-market timing caveat.

Premium reference. The latest valid traded closing premium is compared with valid observations from the preceding 250 market sessions, excluding that latest observation. At least 60 earlier valid premiums are needed for the 5th to 95th percentile reference range. The percentile uses half weight for ties. A reference is descriptive, not a forecast of fair value or an entry signal. Each fund prints its last premium date and age in market sessions. A streak counts consecutive traded sessions above +1% or below −1%; missing NAVs, missing trades and a change of side break it. A stale fund’s streak ends on its printed observation date and is not extended to today.

Tracking consistency. The window is the 36 calendar months ending before the latest data month. Even if the latest session is a month-end, its calendar month is conservatively excluded. The median, middle half, worst observed one-year gap and fraction of observations below zero require at least twelve valid trailing-one-year observations. Overlapping annual returns are not independent trials. A fund with short or missing benchmark history is not given zero tracking difference. The chart preserves missing monthly observations and shows both trailing one-year and annualised three-year NAV gaps. The current active-fund universe is for today’s comparison, not a historical strategy test.

Fees. Strict plan-free scheme-name matches retrieve five years of AMFI observations. Each month keeps its last available observation. Reported total TER before April 2026 and base expenses from April 2026 are separate series; they are never silently joined across the reporting-definition change. Name changes can shorten matched history. An illustrative annual fee is holding amount × expense ratio / 100, assuming the holding amount stays constant. Fees already enter NAV returns and are not added again to tracking difference.

Assets. The available five-year history uses quarterly average AUM for the current AMFI scheme code, rather than summing overlapping old/new scheme codes. This can shorten history around renamings. Quarter-end labels denote the quarter over which assets were averaged. Changes include market-price effects and must not be described as subscriptions or net flows.

Order sizes. The quote table now requires at least five captured sessions for each individual fund. Actual per-fund first/last dates and sample counts are retained. For ₹1 lakh and ₹10 lakh notionals, the published buy/sell capacity is the fraction of all snapshots with valid two-sided quotes and sufficient quantity at the best offer/bid. It does not include lower book levels. The cost illustration is order notional × median quoted spread in bps / 10,000, a buy-and-sell round trip at unchanged quotes, before brokerage and taxes. There is no prediction of exact fills, market impact or slippage; stored five-level aggregate quantities do not contain the price at every level.

Same-index turnover correction. The older same-index table now uses each fund’s summed turnover over the same 60 market sessions, divided by summed group turnover. Its group daily value is the 60-session mean. Ratios of fund medians are no longer presented as trading market shares.

Checks: pipeline/tests/test_etf_insights.py, audits/etf/check_insights.py and audits/etf/check_explorer.mjs. Each profile has a JSON link; chart table/CSV controls follow the selected measure and fund.

This note is the repository file docs/methods/etf.md, rendered as-is.