I have a habit that irritates people at conferences. When someone slides a chart across the table and says "the data shows," I turn it over and look for the footnote. Last month, in a hotel lobby near Termini, a colleague handed me a printout of a BIS working paper and read the headline number aloud: a one basis point increase in deposit rates, and ten weeks later stablecoin growth falls by twelve percent.
I did the arithmetic before the espresso arrived. The Federal Reserve raised 525 basis points between March 2022 and July 2023. If that sentence meant what it seemed to mean, the stablecoin market should have collapsed into negative territory so deep it would have left a crater visible from orbit. It didn't. Supply contracted mildly, wobbled through the USDC depeg in March 2023, and then recovered. So either the paper is wrong or the sentence is wrong โ and after twenty-eight years of watching numbers travel from spreadsheet to press release, I know which one it usually is. A coefficient was ripped out of its units and handed to someone who was told to make it vivid.
That gap โ between what a regression says and what a headline claims โ is the whole of this article. The direction of the BIS finding is probably correct. The magnitude is almost certainly unusable. And the policy apparatus currently being built on top of it does not care about the difference.
This is not a technical paper about protocols. There is no sequencer to audit here, no admin key to check, no upgradeable proxy to trace. It is an empirical monetary economics paper, and its technical content lives entirely in one place: the identification strategy. That is where I want to spend our time together, because in a bull market nobody reads the identification strategy, and in a bear market the identification strategy is the only thing that survives.
Let me set the scene properly, because the setting matters more than the finding.
The Bank for International Settlements published a working paper โ and I want to be precise about that phrase, because the distinction is not academic pedantry. A working paper is not a peer-reviewed publication. It carries an explicit disclaimer stating that the views do not represent those of the BIS or its member central banks. It is a conversation opener dressed in institutional letterhead. I have written things like this myself; in 2023 I circulated a governance report that identified twelve critical centralization risks across three major lending protocols, and I was careful to call it a report and not a standard, because a report invites argument and a standard closes it.
The paper's claim is that deposit rates influence stablecoin flows. That sounds unremarkable until you notice the direction of causation. Most of the literature runs the other way: stablecoins affect bank funding, stablecoins affect monetary policy transmission, stablecoins disintermediate. This paper reverses the arrow. Bank deposit pricing, it argues, is an input to stablecoin demand. The authors describe their contribution as the first empirical evidence of that channel. I would call it a micro-innovation with paradigm-adjacent ambition โ a new viewing angle on a familiar object rather than a new object entirely.
The empirical window runs from 2022 to 2025, roughly 339 weeks of data, deliberately chosen as a high-rate regime. The dependent variable is stablecoin growth measured over rolling ten-week windows. The independent variable is bank savings deposit rates. There are two estimation approaches. The baseline method produces a three percentage point decline in ten-week growth. The instrumental variable method produces nine percentage points. Both are described as evidence for the same claim.
The paper lands in a specific legislative moment, and this is the part that gets lost in the coverage. The GENIUS Act has legally foreclosed the payment of yield by stablecoin issuers. The Federal Reserve, in a September proposal, has moved toward restricting third-party arrangements that would route reserve income back to holders. Circle's distribution arrangement with Coinbase โ a revenue share on reserve interest โ sits directly in the path of that proposal. And the paper's central premise, that stablecoin holders earn zero, is not a market observation but a legal fact. The paper is describing a world that legislation created and that legislation may soon unmake.
That is a strange thing to build a monetary transmission argument on. It is like proving a theorem about the behavior of a river and noting, in a footnote, that the river is a canal.
I spent the last two years in rooms in Rome and Brussels where this exact question โ how to hold a dollar-denominated instrument without generating a security โ was negotiated line by line by lawyers who understood the plumbing better than most engineers I know. What I learned there is that compliance is not a constraint layered on top of a protocol. Compliance is a protocol. And a compliance protocol with a statutory expiration date is a protocol whose upgrade path runs through parliament.
Now the part that matters. Let me walk through the identification strategy with the same suspicion I bring to a contract audit, because the failure modes are structurally similar. In code, a weak invariant produces plausible-looking state transitions that break under adversarial input. In econometrics, a weak instrument produces plausible-looking coefficients that break under a different sample. Same disease, different notation.
The instrumental variable approach relies on the assumption that deposit rate changes are exogenous to stablecoin demand, and it leans on the observation that bank pass-through is larger when the federal funds rate exceeds three percent. That is a reasonable economic story. The problem is the first-stage F-statistic: 9.78.
The conventional threshold for instrument relevance is ten. Nine point seven eight is not a catastrophe. It is worse than a catastrophe in one specific way: it is close enough that a careless reader will wave it through, and far enough that a careful reader cannot. This is the statistical equivalent of a smart contract that passes its test suite with one skipped assertion. The tests are green. The invariant is not proven.
When an instrument is weak, the two-stage estimator becomes biased toward the OLS estimate in finite samples, and its standard errors become unreliable in ways that conventional reporting does not capture. You do not get a wide confidence interval that honestly says "we don't know." You get a confident-looking point estimate that is partly a reflection of the endogenous variable you were trying to instrument in the first place. Weak instruments do not produce noise. They produce false confidence. That is precisely what makes them dangerous in policy settings, because a policymaker reading a table cannot see the F-statistic unless somebody puts it in front of them.
I have watched this exact failure mode play out in DeFi governance. A proposal cites an oracle as decentralized because it has seven reporters. Nobody checks whether those seven reporters are independent. The number looks like a threshold; the threshold looks like a proof. It never is. Chaos is just order waiting to be optimized, and a number that looks tidy is usually the place where the optimization stopped.
Here is the sentence: a one basis point rise in deposit rates produces roughly a twelve percent decline in ten-week growth.
Read it literally. The Fed moved 525 basis points across the sample. Linear extrapolation puts ten-week growth at a figure so negative that stablecoin supply would have to have gone below zero โ an impossibility for a liability that redeems at par. Reality: the aggregate market expanded roughly 44-fold over six and a half years, from something near five billion to something near two hundred twenty billion, an annualized compound rate well above 6.5 percent. Even in the contraction of 2022, the drawdown was measured in tens of percent over quarters, not in the kind of collapse the linear reading implies.
So the sentence cannot be literally true. The most plausible reconstruction โ and I want to flag this as inference rather than assertion, at moderate confidence โ is that "one basis point" is a journalistic compression of a discrete scenario comparison. The real claim is probably: in a rising-rate environment, ten-week growth falls from a mean of about fifteen percent to about twelve percent under the baseline method, or to about six percent under the instrumental variable method. That is a scenario contrast, not an elasticity. It is not "for every basis point, this much." It is "in this regime versus that regime, this much less."
The distinction is enormous and it evaporates in translation. An elasticity is a structural parameter you can plug into a model. A scenario contrast is a description of one historical episode. If you take the second and use it as the first, you will build a supply forecast for USDC that is wrong by an order of magnitude, and you will not find out until it matters.
I have been on the receiving end of this. In 2021 I helped draft a governance whitepaper arguing that smart contracts are a new form of social contract. The most-quoted line from that document was a metaphor I wrote in twenty minutes at the end of a long day, and for two years people cited it as if it were a specification. When a number or a phrase escapes its context, it does not become simpler. It becomes a different claim wearing the same clothes.
Twelve percent versus six percent. Three percentage points versus nine. The baseline method and the instrumental variable method are supposed to be estimating the same underlying quantity, and they differ by a factor of three. In a physics paper, that discrepancy would be the paper. Here it appears as a robustness note.
What does the discrepancy tell us? It tells us the direction is robust and the magnitude is not. Both methods point the same way: higher deposit rates, slower stablecoin growth. That directional agreement across two specifications with different identifying assumptions is genuine evidence, and I do not want to dismiss it. But when two estimators disagree by threefold, at least one of them is substantially wrong, and possibly both are. You cannot average them. You cannot pick the one you like. You can only say: the sign survives, the size does not.
This is what I mean by direction without magnitude. It is a real finding. It is also a finding that cannot support a quantitative policy calibration, and the policy conversation is entirely quantitative. "How much would permitting yield on stablecoins reduce bank deposits?" is a magnitude question. The paper cannot answer it. It can only say: in some direction, somewhat.
For anyone building a model โ a treasury desk, a risk team, a regulator's stress test โ the practical instruction is blunt. Cite this paper for the sign, and go find your own number for the size. Do not import 12 percent. Do not import 6 percent. Do not import 9 percentage points. Those numbers are artifacts of a specification, not properties of the world.
Ten-week rolling growth computed on 339 weeks of data produces a great many overlapping observations. Week one's window and week two's window share nine-tenths of their content. If the standard errors are not corrected for this serial dependence โ Newey-West or an equivalent โ the t-statistics will be systematically overstated. This is not a subtle point; it is the first thing a referee checks in a time-series paper. In a working paper, it may or may not have been done, and the coverage does not tell us.
There is a second, quieter problem with the fifteen percent mean. Ten-week growth averaging fifteen percent, sustained, implies an annualized compounding rate that the actual six-and-a-half-year record does not support. The arithmetic is unambiguous: the mean is being pulled up by the bull-market windows and does not describe a steady state. A mean contaminated by regime is not a parameter. It is a memory of 2021.
Anyone who lived through 2021 knows this in their body. The numbers from that period are not measurements of a system; they are measurements of a mood. I hosted anti-hype workshops through the bear market precisely because I had watched too many people build models on top of mood data and then discover, painfully, that the model had learned the mood and not the mechanism. Two hundred developers came through those rooms. Most of them had never been taught to ask whether a mean was stationary before they built a forecast on top of it.
Here is the finding I keep returning to, the one that should have reshaped the narrative and instead got buried in a robustness table.
The paper's mechanism is deposit substitution: American savers move money between bank accounts and stablecoins depending on relative yield. If that mechanism is the real one, the effect should be concentrated in USDC, the regulated, US-facing, institutionally-adjacent dollar token. USDT, by contrast, has a user base that is substantially offshore and concentrated in emerging markets โ places where the relevant alternative is not a US savings account but a local currency losing twenty percent a year.
And yet USDT also shows a decline, roughly two percentage points.
Think about what that means. If the effect were genuinely about American depositors choosing between a savings account and a token, USDT should be insulated. It isn't. The most parsimonious explanation is that the coefficient is not capturing deposit substitution at all โ it is capturing global dollar liquidity conditions, the Fed cycle, risk appetite, the general cost of holding a dollar-denominated asset when the risk-free rate is five percent. The regression may be measuring the weather and calling it a barometer.
I want to be fair here, because there is a defensible alternative. Offshore dollar holders also face a rising opportunity cost when US rates rise, even without access to a US savings account, because the dollar funding market is global and the offshore dollar is priced off the onshore dollar. So the USDT result is not dispositive. But it shifts the burden of proof. If the paper wants to claim deposit substitution specifically, it needs to explain why the supposedly non-substituting token moves too. It does not.
The authors do attempt a placebo test, and I want to take it seriously before I explain why it underperforms. Their argument is that the second method fails to predict other segments of the crypto market, which they read as evidence that the coefficient is not simply capturing the general crypto cycle.
But think about the mechanics. Stablecoin supply is mechanically bound to crypto trading volume and leverage in a way that other crypto asset prices are not. When leverage expands, you need more dollar-denominated collateral; when it contracts, you redeem. That relationship operates on a timescale of days, not ten-week windows. Other crypto assets โ a governance token, a Layer 1, a mid-cap โ may not respond to the same window at all, because their pricing is driven by narrative, flow, and idiosyncratic supply schedules.
So the placebo test excludes one channel and leaves another wide open: the demand channel. Rates rise, risk appetite falls, leverage contracts, stablecoin demand falls. That is a competing explanation for the exact same coefficient, and it is not addressed. The paper's preferred story is deposit substitution; the alternative is risk-appetite transmission through leverage. Both predict the sign. Neither is excluded by the placebo as constructed.
This is the kind of thing that keeps me up at night about empirical work in this industry. The sign is cheap. Almost every plausible mechanism in crypto predicts the same sign, because almost every plausible mechanism runs through the same two variables: leverage and liquidity. What is expensive is separating the mechanisms, and that requires data the paper does not have.
While I am on the subject of unmodeled contamination: two exogenous shocks land squarely inside the 2022 to 2025 identification window. In May 2022, the collapse of UST triggered redemptions across the stablecoin complex, including USDT, which briefly traded below par. In March 2023, USDC depegged when Silicon Valley Bank failed and a portion of its reserve backing was momentarily in doubt, producing a sharp, mechanical contraction in supply that had nothing to do with deposit rates.
If those events are not handled with event dummies or excluded windows, they contaminate the estimates. A depeg is a supply shock of the first order โ precisely the kind of outlier that a ten-week rolling window will smear across two and a half months of observations. The paper's coverage does not indicate whether this was done. I flag it as an open question rather than a flaw, but it is the kind of open question that determines whether a coefficient is a coefficient or a coincidence.
And then there is the limitation the authors themselves acknowledge, which I think is more serious than they present it.
Two limitations are stated: the sample covers a single high-rate period, and the data is aggregated, so it cannot identify which holders are doing the substituting. Both are treated as caveats. I read them as structural. A study that cannot tell you which bank lost which depositor cannot answer the question regulators actually care about. The policy debate is not "do stablecoins respond to rates in the aggregate." It is "will a specific class of institution lose a specific class of funding." Aggregated data cannot reach that question, and no amount of specification searching will get it there.
What would a better study look like? Bank-level deposit panels, matched to regional stablecoin adoption data. Money market fund yields as the primary regressor rather than a proxy. Event dummies for UST and SVB. Newey-West corrections, disclosed. And ideally a natural experiment โ a discrete rate change with clean timing and a clear control group. That study would cost more and take longer, and it would produce a number you could actually use. This one produces a direction you can cite. Those are different products.
I have spent most of this article on econometrics, so let me ground it in the machinery, because the machinery is what makes the economics behave the way it does.
Reserve composition: short-dated Treasuries, repo, bank deposits. When rates rise, issuer revenue rises โ Circle's reserve income is a direct function of the policy rate. Holder revenue stays at zero, by law. That scissors is not a market outcome; it is a regulatory design choice, and it is the single most important fact about the current stablecoin business model. The issuer captures one hundred percent of the monetary tightening that the holder experiences as pure opportunity cost.
Mint and redeem mechanics: permissioned, fiat rails, T+0 to T+1 depending on the corridor. This friction is precisely why the paper's directional finding is "growth slows" rather than "supply reverses." When the opportunity cost rises, marginal holders exit at the edges โ the ones with the least operational friction, the most rate-sensitive treasuries, the most sophisticated desks. The base does not leave. Redemption friction is the hydraulic resistance in the system, and it is why the pipe bends instead of bursting.
Composability: standard ERC-20, widely used as collateral across DeFi. Here the transmission channel gets interesting and, I think, underexplored. A contraction in stablecoin supply does not just reduce the quantity of dollars; it compresses USD-denominated TVL across lending markets, which raises collateral scarcity, which raises borrowing rates, which further suppresses leverage. The stablecoin supply is not a passive variable in DeFi. It is the base layer of the denominator.
I have written before that Uniswap V4's hooks turn the DEX into programmable Lego and that the complexity spike will scare off most developers. The same dynamic applies to stablecoin integration: the more composable the dollar becomes, the more the entire stack inherits the dollar's monetary policy sensitivity. A yield-bearing wrapper, a lending market, a perpetual venue โ each layer amplifies the base-layer rate signal. When the Fed moves, the whole tower rings.
And this is why I keep insisting that distribution beats technical elegance. The OP Stack and the ZK Stack debate is not really about proving systems; it is about who convinces more projects to deploy chains first. The same is true of dollar rails. Cosmos's IBC is technically beautiful and the application ecosystem remains fragmented, and ATOM captures almost nothing of the value that flows through it. Technical elegance without distribution is a museum piece. The dollar token that wins will not be the most elegant one; it will be the one embedded in the most balance sheets. That is a monetary fact, not a cryptographic one, and no regression in this paper changes it.
Here is where I stop being a careful reader and start being suspicious, which is a different and less comfortable activity.
The paper's conclusion โ that bank deposit pricing still exerts influence over stablecoin demand โ is a conclusion with a constituency. It supports the argument that banks remain the pricing anchor of the dollar system, that stablecoins are a satellite rather than a replacement, and that therefore the existing regulatory perimeter is broadly correct and needs only modest extension. That is a conclusion the banking sector will cite approvingly. It is also a conclusion regulators can cite when defending the current architecture against proposals to permit yield.
I do not think the authors set out to produce a lobbying document. Working papers are usually honest attempts at honest questions, and the disclaimer at the top is real. But I have spent enough time in Brussels negotiating rooms to know that a finding does not need to be commissioned to be useful. Academic work enters the policy bloodstream through the people who find it convenient, not through the people who wrote it. The citation is the currency, and the currency circulates selectively.
Which brings me to the September Fed proposal. If you want to restrict third-party yield arrangements, the cleanest intellectual foundation is a demonstration that deposit rates still discipline stablecoin growth โ that the banking channel remains intact, that the system is not yet disintermediated, that the existing perimeter holds. The paper provides exactly that foundation, with a disclaimer at the top that almost nobody will read. And the disclaimer is the only thing standing between a working paper and a legislative fact.
Now the deeper contrarian point, the one that makes me uneasy.
The paper's foundational premise is that stablecoin holders earn zero. That premise is legally true and economically artificial. It is a fact about the GENIUS Act, not a fact about dollar demand. And what the paper is really measuring, whether it intends to or not, is the behavior of a market that has been administratively prevented from paying interest. It is a study of a price control.
Price controls generate predictable artifacts. When you forbid a yield, the yield reappears somewhere else โ in points programs, in distribution revenue shares, in token incentives, in the spread between what the issuer earns and what the holder receives. Every one of those channels is a workaround, and every workaround is unstable. The Fed's September proposal is an attempt to close the last of them, and the paper's model treats the closure as a permanent feature of the landscape rather than a temporary regulatory victory.
So here is the contrarian claim: the paper's finding has a shelf life, and the shelf life is set by statute, not by data. The moment a compliant yield-bearing dollar instrument reaches scale โ whether as a tokenized money market fund, a permitted payment instrument with a distribution rebate, or something the market invents before the regulators finish drafting โ the zero-yield premise collapses and the estimated elasticities collapse with it. The paper is not wrong. It is time-stamped.
And the final contrarian note, the one I cannot prove and will therefore state as a suspicion. The authors did not run the money market fund regression. I have already said why I think it matters. Let me say what I think it would show: an elasticity substantially larger than the deposit-rate elasticity, because the MMF yield is the actual opportunity cost and the deposit rate is a shadow of it. If that regression were run and published, the conclusion would shift from "banks retain pricing power" to "stablecoins are shadow money market funds, and the zero-yield rule is the only thing preventing them from competing directly."
That is a materially different policy conclusion, and it is the one that matters. I am not accusing anyone of suppression. I am noting that the missing regression is the one that would have been most interesting, and that its absence is convenient for a particular narrative. In a bull market, convenience is the most reliable filter there is. From hype cycles to hydraulic stability โ but the hydraulics only work if you measure the pressure at the right point in the pipe.
So what do we do with this paper?
Cite it for the direction. Ignore it for the magnitude. Treat the twelve percent and the six percent and the nine percentage points as artifacts of a specification, not as properties of the world. Note the F-statistic of 9.78 every single time you reference it, because the person quoting it to you almost certainly won't. And when somebody tells you the study proves banks still set the price of the dollar, ask them whether the study controlled for money market fund yields, and watch what happens to the conversation.
Then ask the question the paper cannot answer: what happens to this entire framework when a compliant dollar instrument finally pays a yield? The GENIUS Act forbids it today. The market is already building the workarounds. The regulators are already drafting against them. And somewhere in that gap between statute and demand, the next stablecoin regime is being assembled by people who have never read a first-stage F-statistic and never will.
The code is cold, but the community is warm โ and the community is holding dollars for reasons that no regression in this paper has yet measured. We are not just users; we are the protocol. And the protocol's interest rate is about to become a political question.