Key takeaways
- Grayscale's bitcoin trust averaged a 25% discount to net asset value in the years it had no redemption programme. Once it had one, its widest discount was 1.56%.
- The SPDR S&P 500 ETF Trust listed 42 authorised participants for 2025. Only 26 created or redeemed anything, and the largest handled 38.5% of $2.1 trillion.
- The iShares Core U.S. Aggregate Bond ETF listed 43 authorised participants in the year to February 2026. Only 4 transacted, and one did 81.4% of the $21.9 billion.
- In mid-March 2020 the largest investment-grade and high-yield bond ETFs recorded discounts to net asset value beyond 5%, and some high-yield funds went past 9%.
- A US ETF whose premium or discount runs past 2% for 7 straight trading days has to publish a discussion of the causes, and keep it up for a year.
Your ETF sits near net asset value because someone can swap the fund for its contents
You buy an ETF at 11am, and the price on the screen isn't the fund's net asset value — the per-share worth of everything it owns. How far apart can those two numbers get, and what stops them drifting?
A swap stops them. A short list of firms, called authorised participants, hold a signed agreement with the fund that lets them do something you can't. They can hand over a basket of the underlying securities and receive new fund shares, or hand back fund shares and receive the securities. US law defines them tightly. Rule 6c-11 describes an authorised participant as a clearing-agency member "which has a written agreement with the exchange-traded fund or one of its service providers that allows the authorized participant to place orders for the purchase and redemption of creation units."
The rest follows from arithmetic. If the fund trades above what it holds, that firm buys the securities, delivers them, receives shares and sells them. The selling pushes the price down. If the fund trades below, it buys shares cheaply, hands them back, receives the securities and sells those instead. The buying pushes the price up.
Neither trade is charity. Both pay only while the gap is wider than the cost of closing it. That single condition is the whole story of when an ETF tracks its net asset value and when it doesn't.
A creation unit is 50,000 shares, and making or unwinding one costs $3,000
The swap doesn't happen in ones and twos. Take the SPDR S&P 500 ETF Trust (SPY). The Bank for International Settlements dates the first ETF to 1993, tracking the S&P 500 index of 500 US companies; this is that fund. Its January 2026 prospectus says these firms "may do so only in large blocks of 50,000 Units known as 'Creation Units.'"
The fee is small and fixed. SPY's trustee charges "the lesser of $3,000 or 0.10% (10 basis points) of the value of one Creation Unit", and the prospectus adds that the fee "is currently $3,000". That tells you something the document never states directly. If $3,000 is the lesser of the two, one creation unit has to be worth at least $3 million.
So the arbitrage is wholesale, and it carries real friction. A gap has to cover the fee, the dealing spread on 500 underlying stocks, a day of funding, and the risk that prices move before the trade settles. Beneath that threshold nobody bothers, and a small premium or discount simply sits there. It's the same reason the headline fee on a fund is never the whole cost — spreads and frictions show up in the gap between a fund's expense ratio and its actual shortfall.
No authorised participant is obliged to do any of it
This is the part most explanations skip. An authorised participant holds a right, not a duty. The Financial Conduct Authority's 2025 occasional paper on ETF mispricing states it plainly: they "enter into a legal contract with the issuer to obtain the right, instead of the obligation to create/redeem ETFs."
Fund prospectuses say the same thing, in the risk factors, where almost nobody reads them. J.P. Morgan's March 2026 prospectus warns that its fund "has a limited number of intermediaries that act as authorized participants and none of these authorized participants is or will be obligated to engage in creation or redemption transactions." The consequence is spelled out too. If those firms "exit the business or are unable to or choose not to proceed with creation and/or redemption orders", then shares "may trade at a discount to NAV and possibly face trading halts and/or delisting."
The thing holding your fund near its net asset value is a commercial decision, retaken every morning, by firms that owe you nothing.
On the first ETF ever launched, one firm did 38.5% of the work
So how short is that short list? US funds have to name every authorised participant and report how much each one created and redeemed, on an annual form called N-CEN. The filings are public and almost unread.
The chart above counts only the firms that actually transacted. SPY's filing for 2025 names 42 authorised participants. Only 26 of them created or redeemed a single share. Between them they moved $2.1 trillion of stock, gross. One firm, Merrill Lynch Professional Clearing, accounted for 38.5% of that, and the five largest for 78.9%.
Bond funds run thinner. In iShares Trust's filing for the year to February 2026, the iShares Core U.S. Aggregate Bond ETF (AGG) had 43 authorised participants signed up and 4 that traded. One of them, J.P. Morgan Securities, did 81.4% of the $21.9 billion. In the big investment-grade corporate bond fund (LQD), 11 of 43 traded, and the largest was Jane Street Capital, a proprietary trading firm rather than a bank, with 30.2% of $122.9 billion. In the high-yield fund (HYG), 8 traded and the top three did 78.5% between them.
That's the mechanism as it actually runs. It isn't a crowd of competing arbitrageurs. On a bond fund it can be one desk having a busy year, and a second one that could step in if the first stopped.
Grayscale's bitcoin trust puts a price on the mechanism
Here's the cleanest measurement available anywhere, because it holds the asset and the sponsor constant and changes one piece of plumbing.
Grayscale Bitcoin Trust held bitcoin and traded over the counter for years with no ongoing way for anyone to hand shares back. Its own annual report for 2025 sets out what that cost. Between May 2015 and January 2024, the maximum premium over net asset value was 142% and the average premium 37%. The maximum discount was 49% and the average discount 25%. The shares closed below net asset value on 725 separate days.
On 11 January 2024 the trust listed on NYSE Arca as an ETF (GBTC), with authorised participants and a redemption programme. From that date to the end of 2025 the same filing reports a maximum premium of 1.68% and an average premium of 0.06%, against a maximum discount of 1.56% and an average discount of 0.08%. On 31 December 2025 it closed at a discount of 0.07%.
Set the two regimes beside each other. Bitcoin's volatility didn't fall. The sponsor didn't change. What changed is that somebody could finally redeem, and the deviation collapsed from tens of percent to fractions of one. Grayscale's own filing names the cause: before listing, the shares traded at premiums and discounts "in part due to the lack of an ongoing redemption program."
For anyone who held the trust before it listed, the discount was a large part of the investment story, which is a different thing from what the spot bitcoin ETFs did to volatility and correlation after they arrived.
The 2% rule that makes a fund explain itself
Regulators treat a persistent gap as a signal rather than a curiosity. Rule 6c-11 makes every US ETF publish, each business day, its net asset value, its market price and its premium or discount, plus a table of how many days it traded at each over the last calendar year. It also has to publish a median bid-ask spread, computed from the best bid and offer "as of the end of each 10 second interval during each trading day of the last 30 calendar days".
The trigger is the interesting part. If the premium or discount is "greater than 2% for more than seven consecutive trading days", the fund has to post a statement saying so, plus "a discussion of the factors that are reasonably believed to have materially contributed" to it, and leave it up for at least a year. Seven days is a regulator's estimate of how long a gap can last before it stops being noise.
The same rule quietly concedes where the swap strains. A fund holding foreign investments may postpone delivery to a redeeming authorised participant when a local holiday or a slow settlement cycle gets in the way, but "in no event later than 15 days" after the shares are tendered. That clause exists because the underlying market is sometimes shut when yours is open.
What a strained mechanism looks like from outside
March 2020 is the episode everyone cites, and it's the more useful one, because the mechanism existed and gave way anyway. The Bank for International Settlements measured it. In mid-March 2020, it reported, "some of the largest ETFs in both the IG and HY segments" — investment grade and high yield — "recorded NAV discounts in excess of 5%". Dispersion in high yield was wider still, with some funds showing "NAV discounts in excess of 9%".
The BIS tied that partly to dealers pulling back: they "provided less support to corporate bond liquidity, potentially limiting the arbitraging of NAV discounts". Its later study of bond ETF arbitrage put it flatly — in March and April 2020 the gap "was not arbitraged away by APs". The pressure wasn't uniform, either. Investment-grade funds holding bonds with under three years left to run recorded a 6.9% discount against a 5.3% average, which the BIS read as stronger selling pressure, consistent with investors rotating into money market funds.
Equity funds weren't spared. The FCA's paper tracks the iShares Core MSCI World UCITS ETF, a large and liquid global fund, at a mispricing of 5% below net asset value on 13 March 2020.
Scale is what separates this from Grayscale. On BlackRock's own numbers, LQD's average absolute premium or discount to net asset value ran at 0.13% across January and February 2020, and 1.39% across March and April. That's more than a tenfold widening. It is also far smaller than the 25% average discount Grayscale's trust carried with no redemption mechanism at all. A mechanism under strain and no mechanism are different orders of problem.
The FCA paper supplies the mechanism underneath the strain. Holding inventory costs an authorised participant balance-sheet capacity, so each firm sets its own limit. As inventory approaches that limit in stressed conditions, the paper finds, an authorised participant's management actions "can exacerbate mispricing for investors". The firm closing your discount is also running its own book, and the second job can win.
The plumbing can jam without any drama at all. ESMA reported that EU settlement fail rates spiked in the middle of April 2025 for ETFs specifically, above the peaks in other market segments, when trade-policy news drove volumes up. A failed settlement is a direct tax on this arbitrage, because a firm that can't deliver can't finish the swap. None of this applies uniformly across exchange-traded products, either: a physically backed gold product is a debt security with its own redemption terms, which is part of why bullion, gold ETCs and mining shares behave so differently.
The strongest case against reading a discount as a failure
BlackRock runs the funds at the centre of March 2020, and it argued the opposite of the standard reading in a presentation to the SEC's fixed income advisory committee. Its case: net asset value for a bond fund is an estimate, struck once a day from dealer quotes on bonds that mostly didn't trade. The ETF price is a real, actionable number. On that reading a discount is the fund telling the truth about prices before the bond market catches up. BlackRock also tested the claim that authorised participants create and redeem in the wrong direction under stress, and reported that it found "no evidence of wrong way arbitrage even in times of market stress."
Both readings can hold at once, and refereeing them matters. If net asset value is stale, the discount is information, and it closes on its own as the underlying reprices. If arbitrage is constrained, the discount is a cost, and it closes when balance sheets free up. The distinction is real for a regulator and for an academic. For someone who sold that morning it isn't, because either way they received the market price and not net asset value.
What these filings cannot tell you
The N-CEN counts are annual, gross and blunt. They record the value each firm created and redeemed across a whole year. They say nothing about who quoted the tightest spread on any given day, nothing about hedge funds that rent an authorised participant's pipe rather than being one, and nothing about how close the fund actually traded to net asset value. A fund with 4 active participants may have been priced faithfully all year long.
They are also a single year, from US filings only. A UCITS ETF listed in London files no N-CEN, so this evidence can't produce the same count for a European fund. The FCA's researchers had to obtain primary market data directly from two ETF sponsors to study the question at all, which tells you how little of it is public.
The Grayscale comparison is clean but it isn't a controlled experiment. The trust's discount before 2024 also reflected a high management fee, a running bet on whether a spot bitcoin ETF would ever be approved, and holders with no other way out. Redemption was the largest change, not the only one. Two years of post-listing data is also a short sample in an asset that has had far rougher years than these.
What would change the conclusion
Three things would, and all three are observable.
If the count of firms actually creating and redeeming a fund fell toward one, the arbitrage would stop being competitive, and the fund's pricing would depend on a single desk's appetite. The N-CEN filings are annual, so that shift shows up late, but it does show up.
If a fund's own premium and discount disclosure started carrying that 2% statement, the fund would be telling you directly that the gap had stopped being noise. That is a deliberately high threshold, so a fund crossing it is publishing an admission rather than a footnote.
And if the underlying market closed — a foreign exchange suspended, a bond market that stops quoting — the swap can't be priced at all. The 15-day postponement buried in Rule 6c-11 becomes the live constraint rather than a footnote, and the price on your screen is whatever the last buyer would pay.
The figure worth watching isn't the fund's performance. It's the distance between the price you paid and what the fund held that day. LedgerTouch stores the price of the trade you actually made rather than the fund's official value, which is where a persistent gap becomes visible. Neither number forecasts anything. Together they tell you what the mechanism was doing on the day you traded.