There is a number in the new Bank for International Settlements working paper on deposit rates and stablecoin growth that cannot be true, and the fact that almost nobody has said so out loud is the most interesting thing about the paper.
The number is this: a one-basis-point increase in bank deposit rates is associated with roughly a twelve percent reduction in stablecoin growth over the following ten weeks.
Read that again, slowly, the way an auditor reads a reconciliation. One basis point. One hundredth of one percent. Twelve percent of a growth rate. The elasticity implied by that sentence is not merely large. It is thermodynamically strange. It implies that a deposit rate drifting from 0.05% to 0.06% rearranges a meaningful fraction of the dollar-denominated shadow banking system โ that a rounding error in a savings account statement can move billions of dollars of offshore liquidity.
Now run the arithmetic forward, because this is what I do. I hunt the story that the chart hides. Between March 2022 and July 2023 the Federal Reserve raised its policy rate by 525 basis points. If the paper's headline coefficient were a linear elasticity, ten-week stablecoin growth would not have "slowed." It would have collapsed so deeply negative that the aggregate supply of USDC and USDT should have inverted itself into a negative number โ a stablecoin market capitalization of less than zero dollars. What actually happened is that combined stablecoin supply contracted modestly from its peak, held, and then resumed growing.
So one of three things is true. Either the BIS has published a paper containing a catastrophic arithmetic error. Or the number everyone is quoting is being read in a way the authors never intended. Or โ and this is where the forensic work starts โ the number is doing rhetorical labor that the regression underneath it cannot support.
Tracing the ghost in the code, in this case, means tracing it in a table of coefficients, because this paper has no code. There is no protocol, no contract, no sequencer, no admin key, no proxy upgrade path. What it has instead is an identification strategy, and an identification strategy is the closest thing monetary economics has to a smart contract: a set of pre-committed rules that determine which conclusions are permitted to execute and which are supposed to revert.
Let me walk you through what actually executes here โ and what silently reverts.
The Paper, Stripped to Its Load-Bearing Beams
Start with what the document is, because in this market, provenance is the first thing that gets laundered.
This is a BIS working paper. Not a journal article, not a policy note, not a staff report with institutional sign-off. A working paper is the economic equivalent of a testnet deployment: it is public, it is citable, it is frequently excellent, and it carries a disclaimer stating that the views do not represent the BIS or any of its member central banks. That disclaimer is not boilerplate. It is a legal shield, and it is being ignored by nearly every person who has shared the headline number this week.
The paper's contribution, as the authors frame it, is a reversal of causal direction. The existing literature on stablecoins and monetary policy has overwhelmingly studied one arrow: stablecoins affecting banks, affecting money market funds, affecting the transmission of policy through the deposit channel. This paper points the arrow the other way โ bank deposit pricing affecting stablecoin flows. That is a genuinely new angle, and I want to be fair to it, because the instinct behind the paper is correct. Stablecoins are not an independent monetary universe. They are a dollar instrument, and dollar instruments are priced off the dollar's risk-free curve.
So the innovation is real. The question is whether the execution carries it.
The dataset spans roughly 339 weeks. Growth is measured with a ten-week rolling window. The core finding is presented in two flavors. Under the baseline specification, rising deposit rates are associated with ten-week stablecoin growth falling from a mean of about 15% to about 12% โ a three-percentage-point decline. Under an instrumental-variables specification, the same shock produces a decline to about 6% โ a nine-percentage-point decline. The instrument is built on the observation that when the federal funds rate sits above 3%, banks pass through a larger share of rate changes to depositors, making the deposit rate more responsive and therefore more usable as an exogenous shifter.
And there is a placebo test: the second method fails to predict the performance of other segments of the crypto market, which the authors offer as evidence that they are capturing a stablecoin-specific deposit-substitution channel rather than a general crypto risk cycle.
That is the skeleton. Now let me start pulling on the joints, because three of them are loose and one of them is load-bearing in a way that should worry anyone building a model on top of this.
The Identification Strategy Is the Real Technical Surface
I want to be precise about something that gets lost in the discourse: this paper's technical content is not its subject matter. It is its econometrics. There is no cryptographic primitive to audit here. There is no circuit, no proving system, no consensus mechanism. The entire technical edifice is the question of whether the variation being exploited is actually exogenous โ whether the thing the authors call a cause is a cause, or merely a correlate wearing a lab coat.
I have spent enough time on both sides of this fence to believe the analogy is exact. When I was twenty-one, still an undergraduate in Doha, I spent weeks inside the Tezos whitepaper not because I cared about the token but because the formal verification claims were unusual and I wanted to know whether they were real. They mostly were, which is why that piece got five thousand reads from an unknown author. The lesson I took from it โ and the lesson that has held for fourteen years of watching this industry โ is that credibility lives in the verification layer, not the marketing layer. A project's whitepaper can say anything. Its proof obligations are what constrain it.
An identification strategy is a proof obligation. It says: given these assumptions, this conclusion follows. Break an assumption, and the conclusion does not become "approximately true." It becomes unsupported.
So let me break them in order of severity.
The One-Basis-Point Ghost
The first problem is dimensional, and it is the kind of problem that a careful reader catches in ninety seconds and a careless one repeats for a year.
The reported coefficient describes a one-basis-point change in deposit rates producing a twelve percent swing in ten-week growth. If that is a linear elasticity โ the natural reading of the sentence as constructed โ then the implied effect of the actual 2022โ2023 hiking cycle is absurd, as I laid out at the top. The stablecoin market would have to have gone negative. It did not. Therefore the sentence, as written, cannot mean what it appears to mean.
The most charitable and most likely reconstruction is that the "one basis point" is a unit simplification introduced in the reporting layer, and the underlying result is a discrete scenario comparison: in a rising-rate environment, ten-week growth runs at roughly 12% under the baseline method or 6% under the instrumental-variables method, against an unconditional mean of roughly 15%. That reading is internally consistent. It is also a completely different claim. A scenario comparison is not an elasticity. It cannot be multiplied. It cannot be plugged into a supply model. It cannot be used to forecast USDC issuance sensitivity to the next 25 basis points.
This matters enormously, because the thing that travels through the news cycle is the number, and the number is the least portable part of the finding. The headline gets quoted in a bank research note. The bank research note gets quoted in a policy comment letter. The policy comment letter gets read by a staffer who does not have time to reconstruct the specification. And somewhere down that chain, "growth slows by three percentage points in a high-rate scenario" becomes "stablecoins are highly rate-sensitive, therefore deposit pricing is a powerful policy lever."
I have watched this exact laundering pattern before, and it always runs through the same three stations. In 2020, during the liquidity mining summer, I was a junior analyst tracking Aave, Compound, Yearn, and Maker simultaneously, and I noticed that governance participation correlated with token price stability. I wrote it up as a "governance premium" and it went wide. Within a month, people were quoting my thread as evidence that governance tokens were defensive assets. They were not. I had found a correlation in a bull market and the market had converted it into a law of nature. I learned then that the distance between a finding and its citation is where most of the damage happens, and I have been suspicious of clean numbers ever since.
The Threefold Gap: Direction Survives, Magnitude Does Not
Here is the second and most consequential technical judgment I will make in this piece.
The paper's two specifications disagree by a factor of three. Baseline says three percentage points. Instrumental variables says nine. Those are not two estimates of the same quantity converging on a truth. Those are two estimates of the same quantity diverging by 200%, which means at least one of them is substantially wrong, and the paper does not adjudicate between them.
Now, there is a standard econometric defense available here: when the baseline and the IV disagree, the IV is often preferred because it corrects for attenuation bias from measurement error in the endogenous regressor. Deposit rates are indeed measured with error, and the OLS estimate is plausibly too small. That defense is legitimate in general. But it only holds if the instrument is strong, and this instrument is not strong. The first-stage F-statistic is 9.78, which is below the conventional threshold of 10 โ the threshold that exists precisely to flag the regime where weak-instrument bias can push the IV estimate arbitrarily far from the truth, in either direction, with standard errors that lie about their own precision.
Put those two facts next to each other. The specification that produces the large number is the specification with the weak instrument. The specification with the credible instrument strength produces the small number. The paper's headline leans on the larger one, and the larger one is the one that the diagnostic statistics actively distrust.
What does this leave us with? A defensible qualitative claim and an indefensible quantitative one. I am willing to accept, on this evidence, that rising deposit rates are associated with slower stablecoin growth. I am not willing to accept any specific magnitude, and neither should anyone building a model. If you take the nine-percentage-point figure and calibrate a USDC supply elasticity to it, you are building on sand that the authors themselves have flagged as sand. The honest summary is: the direction is robust, the magnitude is unusable, and the two are being reported as if they were the same finding.
Overlapping Windows and the Inflation of Significance
The third problem is mechanical, and it is the one I would flag first in a code review.
Growth is measured over ten-week rolling windows across 339 weeks of data. Rolling windows overlap. Ten-week growth measured at week t and ten-week growth measured at week t+1 share nine of their ten component weeks. This creates severe serial correlation in the dependent variable by construction. Every observation is partly the same observation as its neighbors.
If the standard errors are not corrected for this โ with a Newey-West-type estimator or an equivalent HAC correction with a bandwidth matched to the overlap โ the t-statistics will be systematically overstated. Not slightly. Systematically, and often by a large multiple, because the effective sample size is closer to 34 independent ten-week blocks than to 339 independent observations. The paper may well apply the correction; I cannot verify from the reported figures alone. But the burden here runs the other way. When a paper's headline result is a modest three-percentage-point difference, and its data structure mechanically inflates significance, the reader is entitled to see the corrected standard errors front and center rather than inferring their existence.
There is a second, subtler artifact buried in the descriptive statistics that I want to name, because it reveals something about the sample period. The unconditional mean ten-week growth is about 15%. Compounded over 6.5 years, that rate implies growth far in excess of what actually occurred. Stablecoin supply went from roughly five billion dollars to roughly two hundred and twenty billion over that span โ about forty-four times, which is a compound annual rate in the low thirties, not the sixties-plus that a 15% ten-week average would imply.
The reconciliation is that the mean is dominated by the 2020โ2021 window, when stablecoin supply genuinely did explode. The mean is not a steady-state parameter. It is a bull-market artifact. And a benchmark that is itself an artifact makes every "deviation from the mean" interpretation suspect, because you are measuring distance from a number that describes a period which will not recur in the same form.
I have been on the wrong side of this exact mistake. In early 2022 I was holding Luna, and I was holding it because I had built a mental model of "normal" stablecoin and yield-asset behavior out of a sample that consisted entirely of the most anomalous liquidity environment in the asset class's history. When the environment changed, my model did not degrade gracefully. It inverted. The ten-thousand-word post-mortem I wrote afterward was not really about UST's mechanism, which was legible. It was about trust accounting โ the psychology of a market that had calibrated its sense of normal to a regime that had already ended. This BIS paper has the same structural exposure. Its baseline is a bull-market mean, and its identifying variation comes from the one period in the sample when that mean stopped applying.
The Weak Pulse of the Instrument
The instrument deserves its own section because it is where the paper's ambition lives.
The design exploits the fact that banks pass through rate changes to depositors more aggressively when the policy rate is above 3%. The logic is intuitive: in a high-rate environment, deposit competition intensifies, so a given policy move translates into a larger deposit-rate move. That gives the authors a shifter that is plausibly exogenous to stablecoin demand specifically.
There are two problems. The first is strength: F = 9.78 is a weak pulse. In the weak-instrument regime, the IV estimator is biased toward the OLS estimator in finite samples but can also wander far in either direction depending on the correlation between instrument and error, and the conventional confidence intervals no longer cover at the stated rate. The second problem is subtler and more damaging: the instrument is not really a cross-sectional shifter. It is a time-series indicator of "we are in a high-rate regime." And a high-rate regime is not a clean treatment. It is a bundle that includes tighter financial conditions, lower risk appetite, reduced leverage demand, compressed crypto trading volumes, and a stronger dollar.
Which brings us to the placebo test, and to the reason I think the paper's central interpretive claim is shakier than its statistical claim.
The USDT Tell
Here is the anomaly that I cannot get past, and I think it is the most important single data point in the entire paper.
USDC is the American stablecoin. Its users skew toward US institutions, US-adjacent fintechs, US-regulated counterparties. If the mechanism the authors propose is real โ US depositors switching between bank deposits and dollar tokens based on relative yield โ then USDC is exactly the asset that should respond to US deposit rates.
USDT is the offshore stablecoin. Its user base skews toward emerging markets, offshore exchanges, and jurisdictions where dollar access is the whole point and where a US savings account rate is not remotely the relevant opportunity cost. If the mechanism is deposit substitution, USDT should be substantially insulated from US deposit pricing.
It is not. USDT's growth also declined, by roughly two percentage points.
Sit with that. The asset whose holders cannot access US bank deposits responded to US deposit rates almost as strongly as the asset whose holders can. That is not what a deposit-substitution channel predicts. It is exactly what a global dollar-liquidity channel predicts. When the Fed tightens, dollar funding gets scarce, leverage demand falls, crypto trading volumes compress, and demand for dollar settlement instruments of all kinds softens โ onshore and offshore alike. The coefficient is not measuring "Americans moving money from Chase to USDC." It is measuring the Federal Reserve cycle propagating through the entire dollar system.
This is not a small interpretive quibble. It changes the policy conclusion entirely. If the channel is deposit substitution, then deposit pricing is a lever, and regulators can pull it. If the channel is global liquidity, then deposit pricing is a symptom, and the lever is monetary policy itself โ which everyone already knew was a lever. The paper's headline implies the first. Its own data implies the second.
Two Shocks the Paper Does Not Name
There is a further contamination issue, and it comes from outside the paper rather than inside it, which means it requires the reader to bring their own memory.
The identification window is 2022 through 2025 โ the high-rate period. That window contains two of the largest exogenous shocks in stablecoin history. In March 2023, USDC depegged after the collapse of Silicon Valley Bank exposed its reserve custody arrangements, and supply contracted sharply. In May 2022, the UST collapse triggered redemption pressure across the entire stablecoin complex, including USDT, which briefly traded at a discount.
Both events sit inside the estimation window. Neither is a rate phenomenon. If they are not absorbed with event dummies or excluded from the sample, they are sitting in the error term, and they are large enough relative to the effect being estimated to move it materially. The paper does not appear to disclose handling for either. I want to be careful here โ I am reasoning from what is reported, and the authors may address it in an appendix โ but the omission is significant enough that it should be stated in the abstract rather than buried.
This is precisely the kind of thing that makes me distrust clean results in messy samples. When I audited three small ERC-20 governance contracts in 2018, the vulnerabilities I found were never in the elegant parts. They were in the interaction between two components that each looked correct in isolation โ a timelock that respected its delay but not its ordering, a quorum check that counted the wrong denominator. Contamination in an econometric sample works the same way. Each shock is individually defensible as "just part of the period." Together, they are a structural break wearing a business-cycle costume.
The Variable Mismatch, Which Is the Deepest Problem
Now the objection I consider most damaging, and the one I would raise first if I were refereeing this paper.
The independent variable is the bank savings deposit rate. The dependent variable is stablecoin supply growth. The implicit story is that households compare the yield on a savings account to the yield on a dollar token and reallocate.
But that is not the comparison anyone actually makes. A stablecoin holder's opportunity cost is not the average US savings deposit rate, which for most of the sample period sat in a range that would embarrass a mattress. The relevant alternative is a Treasury bill or a government money market fund, both of which tracked the policy rate closely and both of which were yielding north of 4% at the peak. A rational allocator choosing between holding USDC at zero percent and holding a T-bill at five percent is not thinking about their savings account at all. They are thinking about the bill.
So what is the deposit rate actually proxying for in this regression? Two possibilities, and they have opposite implications. Either the deposit rate is a noisy proxy for the general level of short-term risk-free yields โ in which case the paper has measured a broad interest-rate-level effect and mislabeled it "bank deposit substitution." Or the deposit rate is capturing something specific about bank funding competition โ in which case the paper needs to show that the effect survives controlling for T-bill and MMF yields directly.
It does not appear to. And that omission is not incidental. It is the omission that determines whether the paper's title describes its finding.
The Missing Control Group and Why It Matters
Here is where I move from "this paper has flaws" to something closer to a claim about intent, and I want to flag my confidence level explicitly: this is inference, not proof.
The paper does not appear to run the obvious horse race โ money market fund yields against stablecoin supply. That regression is the natural control. It is the first thing a skeptical referee would ask for. And my strong suspicion is that it was avoided, because the result would be inconvenient in a specific direction.
If you regress stablecoin supply on MMF yields, you will almost certainly find a large, highly significant, negative coefficient. Money market funds are the closest functional substitute for a yield-bearing dollar token, and the substitution is visible to anyone who watched supply migrate into tokenized Treasury products whenever on-chain yield lagged the bill. Finding that elasticity would establish the correct frame: stablecoins are a shadow money market fund, and their supply is governed by the spread between on-chain dollar utility and the risk-free rate.
That frame is true, and it is also politically inconvenient, because it implies that the binding constraint on stablecoin growth is the zero-yield rule itself, not bank deposit pricing. And the zero-yield rule is a legislative choice, not a market outcome. A paper that says "banks can defend their deposit base by pricing competitively" supports one policy narrative. A paper that says "stablecoin demand is a function of the yield prohibition" supports a very different one โ one in which the growth of the asset class is a regulatory artifact that could be unwound by statute.
I will not claim the authors deliberately avoided the horse race. I will claim that the paper's policy usefulness depends on not running it, and that this is the kind of coincidence that deserves a footnote at minimum.
The Zero-Yield Premise Has an Expiration Date
The paper's mechanism rests on an assumption that stablecoin holders earn nothing. That assumption is currently true, and it is true for a reason that has nothing to do with markets.
Under the GENIUS Act framework, stablecoin issuers are prohibited from paying interest to holders. The reserve income โ generated by short-term Treasuries, repo, and bank deposits backing the token โ accrues to the issuer, minus distribution arrangements with exchanges and partners. The holder gets zero, legally, and the issuer captures the entire spread.
This is a remarkable structure, and I want to dwell on it because it is the actual economic fact that the paper is dancing around. The issuer's profit is the full reserve yield. In a 5% rate environment, a hundred billion dollars of supply generates five billion dollars of gross reserve income annually, flowing to a handful of corporate entities, with a distribution share carved out for the platforms that route the tokens. The holder, who supplied the dollar, receives nothing. The spread is not a market-clearing outcome. It is a statutory allocation.
So when the paper finds that rising deposit rates slow stablecoin growth, it is finding that the statutory allocation becomes more costly to holders as rates rise. That is true, and it is nearly tautological once stated properly. The interesting question is not whether zero-yield instruments lose appeal when the risk-free rate rises. It is why the yield is zero in the first place, and who bears the cost of that rule.
And here is the part that should make anyone who cares about regulatory honesty uncomfortable: the cost is borne almost entirely by the compliant, onshore holder. The offshore holder can access yield-bearing dollar instruments through structures that the onshore retail holder cannot. The prohibition does not eliminate yield-seeking behavior. It pushes it into venues where the rule does not reach, which is precisely the pattern I have watched play out across every compliance regime in this industry. KYC theater works the same way. The compliance cost lands on the honest user who files the paperwork; the person with a few wallet holdings and a jurisdictionally ambiguous address routes around it without breaking stride.
A rule that makes the compliant product strictly worse than the non-compliant substitute is not a consumer protection rule. It is a market-share transfer.
What the Stablecoin Stack Actually Is
Since the paper's technical surface is econometric, its actual technical base gets almost no attention, which is a shame, because the mechanics explain one of its more interesting findings.
A dollar stablecoin is a permissioned issuance system on top of a permissionless settlement layer. Minting and redemption run through fiat rails, typically same-day or next-day, subject to banking hours and counterparty checks. The token itself is a standard ERC-20, composable across DeFi, used as collateral, as a quote asset, as the unit of account for lending markets.
That structure explains why the paper finds growth slowing rather than reversing. Redemption is friction-laden. It requires a banking relationship, a counterparty, and time. A holder facing a rising risk-free rate does not instantly exit; they stop adding. The flow response is asymmetric โ new minting is elastic, redemption is sticky. That asymmetry alone can generate the pattern the paper observes without any deposit-substitution channel at all. Growth decelerates because the marginal minter pauses, not because existing holders flee.
The reserve side matters too. Issuer reserves sit in short-term government paper and bank deposits. When rates rise, issuer revenue rises mechanically while holder yield stays at zero. The scissors open. That is a real transfer from holders to issuers, and it is largest exactly in the high-rate window the paper studies. The paper treats this as background. It is arguably the whole story.
And the composability side has second-order consequences that the paper ignores entirely. Stablecoin supply is the denominator of DeFi's dollar-denominated TVL. When supply contracts, dollar-denominated metrics compress for reasons that have nothing to do with asset prices. Every dashboard that tracks "total value locked" is partly tracking stablecoin issuance. A slowdown in stablecoin growth mechanically depresses a dozen metrics that analysts then interpret as risk-off sentiment, when in fact they are watching a supply-side accounting effect. I have made this mistake in my own modeling, and I have watched smarter people than me make it in public.
There is a settlement-cost dimension here as well, and it is the one I would bet on next. The rollup layer that most stablecoin flow now settles through is running on borrowed time. Blob space post-Dencun was priced as if it were infinite, and it is not. When that capacity saturates โ and my working estimate is within two years โ rollup fees reprice upward, and every application whose economics depend on sub-cent settlement inherits a new cost floor. Stablecoin payment corridors are exactly that kind of application. The paper's model has no term for settlement cost. That is a variable that is currently near zero and will not stay there.
The Contrarian Read: This Is a Lobbying Document With Standard Errors
Here is the angle I have not seen anyone take, and I think it is the correct one.
The interesting question about this paper is not whether its econometrics are sound. They are mediocre in specific, nameable ways: a weak instrument, a threefold specification gap, overlapping windows, a proxy variable that may be measuring the wrong price, an unaddressed contamination problem, and a placebo test that does not actually rule out the competing channel. None of those are fatal individually. Together they mean the paper supports a qualitative claim and nothing more.
The interesting question is why this paper, with these limitations, is being amplified right now, in this particular regulatory moment.
Consider the constellation. The GENIUS Act has codified the zero-yield rule. The Federal Reserve has floated a proposal widely understood to target third-party yield arrangements โ the workaround where a platform, rather than the issuer, pays holders for holding. That proposal is the live front in the yield war. And into that moment arrives a BIS working paper whose central finding is that bank deposit pricing still exerts meaningful influence over stablecoin flows.
Who benefits from that finding? Banks, first. If deposit pricing is a live lever, then banks can defend their funding base competitively rather than needing protection. That is a friendlier framing than "stablecoins are draining deposits and we need legislative relief." Second, regulators, who get an academic citation supporting the proposition that the existing monetary architecture retains its grip. The paper is quotable by both sides of the bankโfintech divide, which is the surest sign that its framing was chosen with care.
And who loses? The party whose elasticity is never estimated: money market funds as the true substitute, and holders as the party whose statutory zero is the actual mechanism. The paper is careful, technical, and entirely silent on the question of why the yield is zero. That silence is structural. It has to be, because the paper's policy usefulness depends on it.
The narrative didn't arrive by accident. It arrived on schedule, six months after the regulatory clarity that made it useful โ which is exactly the lag I found when I interviewed fifty traditional finance executives in 2024 and distilled their risk concerns into readiness reports. Narrative adoption trails regulatory clarity by roughly two quarters, consistently, because narratives need a legal fact to attach to before they can propagate. The legal fact here is the yield prohibition. The narrative is that banks remain the pricing authority for the dollar.
The Two Limitations That Are Actually Fatal
The paper discloses two limitations, and I want to be explicit that I do not think they are the cosmetic kind that authors list to seem rigorous.
The first is that the sample covers a single high-rate regime. This is not a robustness footnote. It means the paper has observed one treatment condition and is generalizing to a function. We do not know what the relationship looks like when rates fall, because the sample contains almost no falling-rate variation. The instrument โ built on the policy rate exceeding 3% โ is definitionally inert outside that regime. The paper is a snapshot of a regime, presented as a relationship.
The second is that the data are aggregate. The paper cannot see whose deposits moved, from which banks, in which jurisdictions, at what ticket sizes. It observes a national deposit rate and a global stablecoin supply and infers a household-level substitution between them. That is a very long inferential leap across a very large aggregation gap. And it is precisely the question regulators care about โ not "did stablecoin growth slow" but "which depositors left which institutions." The paper cannot answer that. Its own authors say so. Everyone quoting it has skipped that sentence.
This is where I would push hardest if I were in the room. Aggregate regressions of this shape are useful for establishing that a phenomenon exists. They are nearly useless for establishing its magnitude or its mechanism. And mechanism is the entire policy question here. If the mechanism is household substitution, deposit pricing is a lever and banks have a defense. If the mechanism is global dollar liquidity, deposit pricing is a symptom and the lever is the Fed. If the mechanism is the statutory zero, then the lever is Congress. Three completely different policy conclusions, one regression that cannot distinguish between them.
What I Would Actually Want to See
Since criticism without specification is just noise, here is the study I would fund.
One: run the horse race. Regress stablecoin supply growth on money market fund yields, T-bill yields, and bank deposit rates simultaneously. Report the elasticities side by side. My strong prior is that the MMF coefficient dominates and the deposit coefficient shrinks toward insignificance. If that is what happens, the paper's framing is wrong and we all learn something. If the deposit coefficient survives, the paper's framing is vindicated and I will say so publicly.
Two: split the dependent variable by issuer and by chain, and by redemption behavior. USDC and USDT have different holder bases and different redemption mechanics. If the effect concentrates in USDC and vanishes in USDT, the deposit-substitution story strengthens. If it appears in both, the liquidity story strengthens. The paper already has the USDT result and it points the wrong way, which makes this the single highest-value extension.
Three: build the event dummy set. SVB, UST, and the 2023 banking stress should be explicit controls, not absorbed into the residual. Any specification that does not separate a bank-run shock from a rate shock is estimating a blend of the two.
Four: report corrected standard errors prominently, with the effective sample size stated. If the result survives an honest HAC correction on 34 effective blocks, that is a much stronger paper than one that survives 339 pseudo-independent observations.
Five: publish the replication package. This is a working paper with a headline number that has already been misquoted in the press, in a dimensional direction that a referee should have caught. The fix is not a retraction. The fix is code and data, so that anyone quoting the number can see what the number is.
That last point is not a technicality. It is the whole thing. In an industry where I have spent years tracing ghosts through bytecode and governance contracts, the single most reliable predictor of whether a claim survives scrutiny is whether the claimant showed their work. Papers that publish replication packages tend to be right more often than papers that do not, and the causal direction runs both ways.
What This Means If You Are Holding Dollars On-Chain
Let me translate all of this into something useful, because the audience for this piece is not the BIS.
If you are holding stablecoins right now, in this market, the paper's actual finding is less important than the thing it accidentally documents. Your opportunity cost is the risk-free rate, and you are paying it in full, by statute. Every basis point the Fed holds at the front end is a basis point transferred from you to an issuer balance sheet. That is not a scandal. It is a disclosed structure. But it is the structure, and any analysis that treats stablecoin demand as a function of bank deposit rates rather than the T-bill rate is analyzing the wrong spread.
Practically: the yield prohibition is the binding constraint on your returns, and it is a legislative artifact with a live political fight attached to it. The Fed's proposal targeting third-party yield is the front line. Watch that, not the deposit rate. The deposit rate is a symptom of the rate cycle. The yield rule is the thing that determines whether stablecoins remain a zero-yield instrument or become a competitive one.
And watch the offshore gap. As long as the onshore compliant product pays zero and the offshore structure pays something, supply will migrate along that seam, and the onshore metric will understate real demand. That is not a prediction about price. It is a prediction about where the plumbing goes.
Mining for meaning in a sea of volatility
Here is what I think is actually happening, stripped of the econometrics.
A working paper with a weak instrument, a threefold specification gap, and a headline number that cannot survive dimensional scrutiny has been converted, in under a week, into a widely repeated claim that banks remain the pricing authority for the dollar. That conversion did not happen because the econometrics supported it. It happened because the claim was useful to parties who needed it to be true, at a moment when they needed it.
The paper's qualitative finding is probably right and definitely unremarkable: when the risk-free rate rises and your instrument pays zero, demand for your instrument softens. You did not need 339 weeks of data to learn that. What you need 339 weeks of data for is to make the claim look like a discovery rather than an implication, and to attach a magnitude to it that can be cited in a comment letter.
That is the ghost I keep finding, in code and in regressions alike. The mechanism is usually boring. The packaging is where the intent lives. A stablecoin is a money market fund that is legally forbidden from being one. A working paper is a policy instrument that is legally forbidden from being one. Both disclaimers are on the first page, and both get skipped.
The question I would leave you with is not whether deposit rates affect stablecoin growth. They probably do, a little, in a way that is mostly the Fed cycle wearing a bank's name tag. The question is this: when the zero-yield rule eventually breaks โ and it will break, because a statutory prohibition on paying interest on a hundred-billion-dollar deposit base is not a stable equilibrium in a competitive market โ what happens to a research literature that spent its high-rate years measuring the elasticity of a rule rather than the elasticity of a market?