A Basis Point Too Far: Auditing the Demand-Side Case Against Stablecoin Growth

There is a sentence buried in a recent Bank for International Settlements working paper that should have stopped every stablecoin treasury desk in the world, and instead passed through the news cycle almost untouched. It states, in effect, that a one basis point increase in bank deposit rates is associated with roughly a three percentage point decline in ten-week stablecoin growth — and that the instrumental-variable estimate is approximately triple the ordinary least squares result, putting the figure nearer nine percentage points per basis point. Read that again, slowly, the way you would read a line of Solidity that looks almost right. A basis point is one hundredth of one percent. Three percentage points is three hundred basis points. The coefficient is claiming a multiplier of three hundred to one, and the two-stage version claims nine hundred to one. When a metric fails its own arithmetic before it fails its economics, you are no longer looking at a finding. You are looking at a transcription error wearing the clothes of a finding.

I have spent enough of my career listening to the errors that the metrics ignore to know that the interesting part of this story is not the coefficient. It is what the coefficient is doing to the industry's models, and what the industry's models are doing to capital allocation, and what a statutory deadline of January 18, 2027 is about to do to all of it. Because underneath the bad number sits a genuinely important structural claim — that stablecoin growth is demand-constrained by the price of bank deposits, not supply-constrained by the availability of rails — and underneath that claim sits a piece of legislation that has quietly hard-coded the yield of the world's most important dollar token to zero. The paper's math is broken. Its direction may not be. Distinguishing the two is the entire job.

The context that the headline omitted

The research in question is a BIS working paper by Rashad Ahmed and Iñaki Aldasoro, and its central argument is that deposit rates function as a kind of stubborn gravity on stablecoin adoption. The framing is deliberately unglamorous, and that is precisely what makes it worth reading. For the better part of two years, the industry has narrated its own progress in the language of supply: new corridors, new card integrations, new settlement windows, new partnerships with the incumbents who once regarded crypto as a compliance headache. Visa Direct has moved into commercial production. Mastercard's Move program reaches roughly 130 countries. Every quarter brings another press release announcing that the last mile has been paved. The implicit syllogism is that once the rails exist, adoption follows automatically, because the product is obviously better than the alternative.

The paper's contribution is to point out that the syllogism has a missing premise. Adoption does not follow from availability; it follows from relative price. A stablecoin is, in technical terms, a tokenized promissory note against a reserve account holding cash and short-dated Treasuries. The architecture has not materially changed since 2014. There have been improvements in attestation frequency, in redemption plumbing, in chain selection, in the depth of the dealer relationships behind the mint-and-redeem window — but the core object is still an IOU with a bank on the other side of it. What changed in 2024 and 2025 was not the object. It was the distribution layer wrapped around the object. Visa and Mastercard did not make stablecoins more decentralized or more trustless. They made them acceptable to merchants and acquirers who never want to touch a wallet, and in doing so they performed a subtle act of reclassification: the stablecoin became a backend settlement rail rather than a front-end currency. The end user may complete an entire purchase without ever knowing that a token moved.

That reclassification matters enormously for how we read the industry's own statistics, and it is where my first real objection begins. The source material cites more than four hundred billion dollars in stablecoin payment volume across the first eight months of a year it identifies as 2026 — a dating problem in itself, since the GENIUS Act was signed in July 2025 and the interest prohibition does not take effect until January 18, 2027, which makes a 2026 publication discussing a 2026 measurement window internally awkward at best. Set the calendar aside. The deeper issue is that no measurement basis is disclosed. Does the figure include wallet-to-wallet transfers? Exchange rebalancing? Market-maker inventory shuffling? Bot traffic on high-frequency arbitrage routes? Anyone who has ever pulled a raw transfer count from an explorer knows that the gap between gross on-chain volume and genuine terminal payment for goods and services can be an order of magnitude or more. Protecting the ledger from the volatility of hype requires, at minimum, that we stop quoting gross transfer volume as though it were retail commerce.

Then there is the statute itself, which is the part of this story with actual teeth. The GENIUS Act establishes a federal framework for payment stablecoins, requires one-to-one backing in high-quality liquid assets, restricts rehypothecation, and — in Section 4(a)(11) — prohibits issuers from paying interest or yield to holders. The prohibition takes effect on January 18, 2027. Note the precise scope: it binds the issuer. It does not bind a third-party platform that chooses to run a rewards program, a cashback scheme, a fee waiver, or a points ledger denominated in something that is not technically a dollar. That asymmetry is the whole regulatory ballgame, and I will come back to it, because it determines whether the next three years of stablecoin competition happen at the issuer layer or somewhere else entirely.

The arithmetic that does not close

Let me do the audit properly, because this is the kind of claim that deserves a line-by-line read rather than a summary.

If a one basis point rise in deposit rates reduces ten-week stablecoin growth by three percentage points, then the Federal Reserve's 2022–2023 tightening cycle — roughly 525 basis points of cumulative increases — should have produced a decline in the ten-week growth rate of approximately 1,575 percentage points. Growth rates are not bounded below in principle, but they are bounded in practice by the fact that a market cannot lose more than all of itself. What actually happened is that total stablecoin market capitalization moved from roughly 180 billion dollars to roughly 125 billion dollars between March 2022 and mid-2023, a drawdown on the order of thirty percent. Those two numbers — 1,575 percentage points of deceleration versus a thirty percent contraction — are not in the same universe. They are not even in the same discipline.

The most charitable reading is that the published coefficient is denominated in basis points rather than percentage points, which would reduce the effect by two orders of magnitude and bring it into the realm of the plausible. A second possibility is that the coefficient corresponds to a 100 basis point move rather than a single basis point, which would produce a similar rescaling. A third, and in my experience the most common, is that the first-stage summary lost the standardisation or log transformation that made the original specification dimensionally coherent. Until someone opens the PDF and checks the units, the honest treatment is to mark the figures as unverified and refuse to propagate them.

I have a specific reason for being pedantic here. In 2017, as a twenty-year-old cybersecurity student in Ho Chi Minh City, I spent three months auditing the ERC-20 contracts behind a then-popular token sale, and the vulnerability I eventually found was an integer overflow in the vesting schedule. The code compiled. The tests passed. The type system was internally consistent. What was wrong was not the arithmetic but the assumption about the domain of the arithmetic — the contract believed a number could grow past the point where its representation could hold it, and the boundary between those two worlds was invisible until someone walked across it deliberately. Econometric coefficients have the same failure mode. A regression can be perfectly specified, correctly estimated, and internally coherent while the units of its output have drifted three orders of magnitude away from the units of the world. The number does not announce its own absurdity. Someone has to check it against reality, and the checking is the work.

The practical consequence is not academic. Total addressable market models for stablecoins are built by multiplying elasticities like this one against rate scenarios. Feed a model a coefficient of three percentage points per basis point and the model will tell you that stablecoin supply collapses to zero the moment the Fed nudges the dial, which is not a forecast but a division by zero dressed as a demand curve. Analysts who ingest this number without rescaling it will produce projections that are not merely wrong but directionally inverted, and they will do so with the full confidence of a citation. The quiet confidence of verified, not just claimed is not a stylistic preference. It is the only defence a research desk has against a well-formatted error.

The instrument variable as an unchecked require()

The second audit point concerns methodology, and here I want to borrow a habit from smart contract review rather than from econometrics.

An instrumental variable is a proxy. You cannot observe the causal effect of deposit rates on stablecoin demand directly, because deposit rates and stablecoin demand are jointly determined by a hundred things at once, so you find a variable that moves deposit rates without moving anything else, and you route the inference through it. The source material indicates the instrument is something like the federal funds rate crossing a three percent threshold. For this to be valid, the instrument must satisfy an exclusion restriction: it must affect stablecoin growth only through its effect on deposit rates, and through no other channel whatsoever.

That restriction does not hold, and it does not hold in a way that is obvious rather than subtle. The federal funds rate crossing three percent simultaneously reprices risk assets, compresses DeFi lending yields, changes the opportunity cost of holding any non-yielding instrument, alters venture funding conditions across the entire crypto sector, and shifts the risk appetite of every marginal allocator in the market. It is not a clean lever that touches deposit rates and nothing else. It is a lever attached to the entire machine. The analogy I would use with a developer is a proxy contract that appears to forward a single call but in fact routes through three implementation addresses; you think you are measuring one function's behaviour and you are measuring the composition of several. The estimate that comes out the other end is not wrong because the regression was run incorrectly. It is wrong because the identification strategy assumed a purity the world does not offer.

I ran into a version of this in 2024, when I was asked to review the custodial arrangements of three major firms for compliance with the post-approval ETF guidance. Two of the three were using threshold signature schemes that had been deprecated for exactly this reason — the cryptography verified correctly, the signatures were valid, the wallets functioned. What failed was the policy layer sitting above the cryptography. Verification and validity are not the same property, and conflating them is the single most reliable way to build a system that passes every test and still fails. The BIS instrument passes the mechanical test. Whether it passes the validity test is a different question, and the answer here is no.

There is a third methodological soft spot worth naming. The dependent variable is described as ten-week growth, and the independent variable is a one basis point change in deposit rates. The dimensional mismatch between those two quantities is severe enough that the specification almost certainly included a standardisation step that the summary dropped. Authors usually disclose this. Summaries usually do not. And readers, who are busy, inherit the number without the transformation that made it meaningful.

To be fair to the researchers, and I do want to be fair, they appear to have been unusually candid about their own limits. The source material notes an explicit acknowledgement that the work establishes correlation rather than causation, and specifically that it does not demonstrate that stablecoin growth actually reduced bank deposits or bank lending. That sentence is the most honest line in the paper and the least likely to survive transmission. A headline reading roughly that banks can slow stablecoin adoption just by raising deposit rates adopts a causal register that the body text explicitly disclaims. Headlines travel at network speed; caveats travel at the speed of the people who read footnotes. When the floor drops, the foundation speaks — and what the foundation says here is that the floor was never load-bearing.

Recasting the token economy: who owns the interest

Here is where the analysis becomes genuinely useful, because fiat-backed stablecoins have almost nothing in common with the token economics the industry usually discusses, and applying the standard template produces nonsense.

There is no vesting schedule. There is no team unlock. There is no emissions curve, no governance token, no treasury diversification program, no buyback-and-burn. Supply is elastic and demand-determined: someone deposits a dollar, a token is minted; someone redeems, a token is burned. The interesting economic question is therefore not who gets tokens. It is who gets the interest on the reserve.

Run that question through the stack and the answer is stark. The reserve — short-dated Treasuries, overnight repo, insured bank deposits — generates a yield that is a direct function of the policy rate. At the issuer layer, essentially all of that yield is retained. At the distribution layer, exchanges and wallets negotiate revenue-share arrangements or run their own reward programs funded out of their own margin. At the holder layer, the statutory answer after January 2027 is zero. The entire monetary economics of the largest dollar token in the world reduces to a three-line allocation table, and the line that says "holder: 0%" is not a market outcome. It is a legislated one.

This is why the sensitivity figures in the paper, whatever their true magnitude, are directionally revealing in a way the aggregate numbers are not. The source material reports a markedly higher rate sensitivity for the most compliance-forward issuer — on the order of several percentage points per basis point — against a materially lower sensitivity for the offshore issuer, roughly half as much. That differential is not noise. It is a user-base signal. The compliance-forward issuer's holders skew toward US institutions, treasuries, and funds for whom the token is functionally a deposit substitute; those holders have a live alternative that pays a real rate and carries deposit insurance. The offshore issuer's holders skew toward jurisdictions where the relevant comparison is not a Treasury bill but a local-currency account with a double-digit nominal rate and a higher real inflation rate, or no reliable banking relationship at all. For that user, the question is not "what does this pay" but "can I get dollars." Availability dominates yield.

Which produces the structural conclusion I keep returning to: the GENIUS Act's interest prohibition is not a neutral consumer-protection measure. It is a mandatory reallocation of reserve yield away from holders and toward issuers and distributors, and its practical effect is to convert stablecoin issuance from a service business into a spread business. Once you see that, the rest of the industry's strategic confusion becomes legible. An issuer that cannot pay its customers is not competing on product. It is competing on the width of its own net interest margin, and it is doing so against an incumbent — the commercial bank — that has deposit insurance, a branch network, a lending book, and a century of regulatory capture.

The double bind

The consequence of that reallocation is a trap with two jaws, and I have not seen it stated cleanly anywhere in the commentary I have read.

When policy rates are high, banks win the store-of-value contest. A depositor weighing a Treasury-backed token paying nothing against an insured account paying a real rate does not need a spreadsheet. Stablecoin growth decelerates, exactly as the paper describes, whatever the true coefficient. But when policy rates fall, the issuer's unit economics compress, because reserve income is essentially the entire revenue line. Growth resumes and margin deteriorates in the same motion, driven by the same exogenous variable. The issuer cannot simultaneously maximise scale and unit revenue, because both are governed by a rate it does not control. That is not a business model with a cyclical overlay. That is a structurally hedged position that has been marketed to public shareholders as a growth story.

There is a further asymmetry worth flagging. The prohibition binds issuers and not distributors. A large exchange can fund a USDC rewards program out of trading revenue, or out of the revenue share it already receives, and thereby retain users that the issuer has no legal mechanism to retain. Circle loses a tool. Coinbase keeps one. Value capture migrates up the stack, away from the balance sheet that carries the reserve risk and toward the interface that owns the customer relationship. If I were modelling Circle's terminal valuation multiple, I would be arguing for a re-rating from financial infrastructure toward regulated utility — lower growth, lower multiple, more predictable, more substitutable. That is not a bearish call on stablecoins. It is a call about where in the stack the rent accrues.

And there is a regulatory gift hiding inside the penalty, which almost nobody has priced. Run the Howey test against a payment stablecoin after Section 4(a)(11) takes effect. Money invested: yes. Common enterprise: weak, since the relationship is custodial or creditor-debtor rather than a joint venture. Expectation of profit: no — and this is the point. The statute has eliminated the prong by fiat. Profits from the efforts of others: no, because value is anchored to the reserve rather than to managerial skill. The interest prohibition, which handicaps the issuer commercially, simultaneously hardens the asset's non-security classification. The industry spent years lobbying for exactly that clarity and has mostly failed to notice that it received it as a side effect of a restriction it opposed. The audit trail, read carefully, is a narrative of trust — including when the entry that protects you looks like the entry that hurts you.

What the paper does not see

The contrarian case against this research is not that its direction is wrong. It is that its frame is too narrow to support the conclusions being drawn from it, and the omissions cluster in a way that suggests a blind spot rather than an oversight.

Start with the largest one. The paper does not discuss tokenized deposits at all. Not as a competitor, not as a control, not as a caveat. This is extraordinary, because a bank-issued deposit token is a strict superset of a payment stablecoin on every dimension the paper cares about. It settles on the same rails. It carries the same programmability. It is backed by bank credit rather than by a bankruptcy-remote trust. It can pay interest, because it is a deposit and deposits pay interest. And in the United States it is eligible for deposit insurance up to the statutory cap of 250,000 dollars per depositor per institution. If JPMorgan, Citi, and HSBC scale deposit tokens into production, the competitive question stops being "can a stablecoin out-yield a savings account" and becomes "why would an institutional treasurer hold the uninsured version." The stablecoin's addressable market then narrows to precisely the population the banking system does not want: the underbanked, the cross-border remitter, the on-chain native, the sanctions-exposed jurisdiction. That is a real market. It is not the total addressable market the current valuation narratives assume. An analysis that claims to model the demand-side constraint on stablecoins while omitting the one product that removes the constraint entirely is not a partial analysis; it is a mis-specified one.

Second, the paper treats the banking sector's reaction function as frictionless, and it is not. Raising deposit rates costs net interest margin. Aggressive deposit gathering invites supervisory scrutiny, particularly for institutions already carrying unrealised losses on their securities portfolios. And deposit betas — the fraction of a market rate change that actually passes through to depositors — are famously partial and lagged. During the 2022–2023 cycle, the realised pass-through was slow and uneven, which is precisely why the deposit outflow episode that did occur was concentrated and dramatic rather than smooth. If deposit betas are well below one, then the gravity the paper describes is weaker and slower than a naive reading suggests, and the stablecoin demand curve is correspondingly flatter. The model assumes banks can raise rates freely, instantly, and without cost. Real banks do all three imperfectly, and the imperfection is the entire competitive opening.

Third, there is a feedback loop the paper's single-direction framing cannot accommodate. Stablecoin reserves are a standing bid for short-dated Treasuries and overnight repo. If the sector grows, that bid grows, which puts downward pressure on short-end yields, which improves the relative attractiveness of stablecoins, which supports further growth. The magnitude is probably small relative to the Treasury market, and I would not want to overstate it, but the sign matters: a model that runs only from rates to stablecoins and never from stablecoins back to rates has a specification error, not a rounding error. Rooted in the past, secure for the future is a fine posture for a settlement layer, but it does require that the model acknowledge the loop it is embedded in.

Fourth, the sample period is a high-rate regime. The evidence runs from 2022 through 2025, which is to say through the sharpest tightening cycle in four decades followed by a plateau. A sustained cutting cycle would weaken the gravitational pull automatically, without any change in stablecoin technology, distribution, or regulation. The paper does not run that scenario, and readers should. The conclusion is regime-dependent in a way the headline is not.

Fifth, and most fundamentally, the paper reduces the competitive contest to yield. That is like evaluating a Layer 2 by transaction throughput alone. I spent two weeks in 2023 reverse-engineering the consensus mechanisms of three major sequencers and quantifying the exact share of block production controlled by single operators, and the lesson I took from that exercise was that the metrics everyone quotes are almost never the metrics that determine outcomes. For stablecoins, the dimensions that banks cannot replicate are not rates. They are atomic settlement, round-the-clock availability including weekends and holidays, programmability at the protocol level, composability with on-chain credit markets, and the capacity for machine-to-machine micropayments at a granularity no card network can economically clear. When the counterparty is an autonomous agent executing a thousand sub-cent transactions an hour, the concept of a deposit rate does not apply, because the agent has no alternative instrument to compare against. The paper measures the one thing banks win on and does not measure the several things they structurally cannot do.

Finally, there is the measurement problem I raised at the outset, and it deserves to be restated as a substantive rather than a technical criticism. The four hundred billion dollar payment figure, whatever its true period, almost certainly includes volume that is not payment. Internal transfers between a user's own wallets. Exchange hot-wallet rotations. Market-maker inventory movements. Liquidity provision on automated market makers, which is not commerce but inventory management. Until the industry publishes a measurement basis that separates terminal settlement from gross transfer volume, every TAM model built on that number is compounding an unknown error with a known one.

The takeaway: what the January 2027 window actually decides

The date to circle is January 18, 2027. Everything about stablecoin economics between now and then is a negotiation over what that date means.

Three things will determine the outcome. First, how the implementing agencies — the OCC and the Treasury — define disguised yield. If a rewards program funded by an affiliated exchange, a fee rebate, a points ledger redeemable for a non-dollar asset, or a yield-bearing wrapper token are all treated as prohibited interest by construction, the prohibition has real force and the distribution-layer arbitrage closes. If the definition is narrow and formalistic, the entire industry will re-engineer around it within two quarters, and the statutory text will have accomplished nothing except a compliance-cost transfer to lawyers. Second, whether the framework permits bank subsidiaries to issue payment stablecoins. If it does, the largest institutions will hold both legs — the interest-bearing insured deposit and the programmability-enabled token — and independent issuers will be competing against a vertically integrated superset. Third, the pace at which tokenized deposits actually reach production, because that is the product that makes the yield question moot rather than merely unfavourable.

What I would watch, concretely: the original BIS working paper for the units on those coefficients, because nothing downstream is usable until that is resolved; the quarterly circulation figures in the compliance-forward issuer's filings, regressed against the effective federal funds rate, because that is the cleanest external test of the sensitivity claim available in public data; the stablecoin share data on any reputable dashboard, split by issuer, to see whether the sensitivity differential persists through a rate plateau; and any rulemaking notice that uses the phrase "substantially similar to interest," because that phrase is where the next three years of competitive advantage will be won or lost.

Guarding the gate, not just the gold, means being precise about what this research actually establishes. It establishes that the industry's supply-side triumphalism is incomplete — that rails do not create demand, and that a token whose statutory yield is zero is competing against an instrument that pays. It does not establish a usable elasticity, a causal mechanism robust to instrument choice, or a forecast. The direction is worth holding. The number is not.

So the question I would leave with the desk is not whether stablecoins can pay interest after January 2027. The statute already answered that. The question is whether a rail that is legally forbidden from paying interest can still become the rail that everything else runs on — and whether the answer changes when the counterparty on the other side of the transaction is not a depositor comparing yields, but a machine that has never opened a savings account and never will.