The market, measured
from its own filings.

Measured from SEC Form N-PORT structured data, 2019 Q4 – 2026 Q2 (27 quarters), and independently audited 2026-07-28 by a separate agent instructed to break it. Every statistic below carries its denominator in the same sentence.

A mid-period exit is not misconduct. These are aggregate descriptions of fund-level filings — not causal claims, not scores, and not advice.

Two choices in this measurement are discretionary enough to move the headline. Both are shown whole, so a reader sees the range before deciding what to make of any single figure.

How wide is the end-of-period window?

Used: reset month + prior month — 19.7% of the observed calendar

The 2-month window is empirically justified — of 276 annual-series ETFs with at least 18 months of history and $50M of flow, 265 (96%) peak exactly on the inferred reset month (121) or in the month before (144) — but the headline ranges from 34.5% to 72.4% across defensible definitions, and a reader is entitled to see that.

How are negative-signed redemption filings treated?

Used: absolute value

The ratio is robust; the dollar total is not. The total is quoted as "$103-160B depending on sign treatment", never as a bare figure. The raw signed sum is the one indefensible option: filers themselves amend these cells, and every one of the 12 amended values is a sign flip in the same direction.

Carried verbatim from the measurement document — the limits ride with the numbers, not behind a link.

• These are fund-level flows, not investor trades. Creations and redemptions happen when an authorized participant assembles or breaks a basket. An investor who sells shares on the exchange to another investor produces no flow at all. This measures money leaving the fund, which understates how often individual holders turn over.
• Monthly granularity. A redemption on 15 December and one on 31 December are indistinguishable, and only the second is unambiguously at term.
• 22 of 471 ETFs are unmeasured because no reset month can be read from the name.
• Boundary sharpness. Innovator U.S. Equity Power Buffer ETF – December has 38% of its redemptions in November (scored at term) and 17% in October (scored mid-period, despite 11 of 12 months having been held). The window is a line drawn through a continuum.
• Quarterly series get a window spanning 8 of 12 months, so "19.7% of the calendar" is a blend across series types, not a constant.
• Two-year series are treated as resetting annually, which slightly over-credits the window.
• No causal claim. A mid-period exit is not misconduct. Investors have every right to sell, and exits can be forced by circumstance. The number says how often the protection the product advertises could not have been delivered — not why.

research/buffer_universe.md, carried verbatim

An earlier headline about the share of "net exits" leaving mid-period failed falsification — randomly permuting each fund's redemptions across its own months, destroying all exit-timing information, left it essentially unchanged, while the gross statistic collapsed toward its null — so it was deleted and is not quoted here or anywhere. Only the gross figure carries signal.

Each month, every fund in the country tells the SEC how much came in and how much went out. This is that data, reduced to the funds investors put the most money into — and pulled the most out of — across the whole market rather than one broker’s customers. Data vintage: 2025-05 to 2026-06.

fixed alphabetical — never ranked, no sort controls

• AllianzIM: 55% left mid-period, of $19.3B of gross redemptions across 46 funds — a share of that shelf's own redemptions, not of the market.
• Calamos: 87% left mid-period, of $1.3B of gross redemptions across 32 funds — a share of that shelf's own redemptions, not of the market.
• First Trust / FT Vest: 34% left mid-period, of $66.6B of gross redemptions across 108 funds — a share of that shelf's own redemptions, not of the market.
• Innovator: 51% left mid-period, of $68.0B of gross redemptions across 173 funds — a share of that shelf's own redemptions, not of the market.
• iShares: 68% left mid-period, of $0.6B of gross redemptions across 8 funds — a share of that shelf's own redemptions, not of the market.
• Pacer Swan: 29% left mid-period, of $1.9B of gross redemptions across 12 funds — a share of that shelf's own redemptions, not of the market.
• PGIM: 19% left mid-period, of $0.6B of gross redemptions across 40 funds — a share of that shelf's own redemptions, not of the market.
• TrueShares: 37% left mid-period, of $1.3B of gross redemptions across 22 funds — a share of that shelf's own redemptions, not of the market.

The spread is the finding. Behaviour is not a property of the product — it is a property of who sold it and how. PGIM at 19% and Calamos at 87% hold structurally similar instruments.

The endpoints of that 19%-87% spread are the two smallest cells — PGIM at $0.6B and Calamos at $1.3B, together 1.2% of the $159.7B measured. The by-issuer table is computed at the single two-month window, while the aggregate headline swings from 34.5% to 72.4% across defensible windows; that sensitivity has not been re-run per issuer, so the issuer gap sits inside an untested method band.

A mid-period exit is not misconduct, and none of this is a score: these are fund-level flow shares from the issuers' own N-PORT filings, shown with their denominators, in alphabetical order.

Figures that replaced overturned copy, from the module of record (figures.py) — each pinned to its source, its date, and its denominator.

Cross-check against the bundled classify output (research/measurement.json): candidates 654, excluded 29 (each with a recorded reason: 29 entries), ladders held out 30, counted 582, ETFs 471, measurable 449 — matches the flagship block above.

Reproducing it python3 fetch_nport.py → python3 universe.py → python3 classify.py — 29 exclusions, each recorded with a reason in research/measurement.json; independently audited 2026-07-28 by a separate agent with no stake in the result, instructed to break it. N-PORT original-vs-amendment preference kept: amendments and latest filings are preferred throughout. Source: https://www.sec.gov/files/dera/data/form-n-port-data-sets/