The Fed made its stress test readable, and made it surprise you less
Publishing the models behind the US stress test is a genuine reform. Damping how much the resulting capital requirement can move is a separate decision, and it travelled in under the same cover.
Large American banks must hold a cushion of shareholders' money to absorb losses. Each year the Federal Reserve, the US central bank, runs a hypothetical severe recession through each bank's books and uses the result to set part of that cushion. It has now finalised rules to publish more about how the test works and to make the resulting requirement move less from year to year.
Publishing the Fed's stress test models is overdue and right. Damping how much the resulting capital requirement can move is a separate decision that mostly benefits the banks, and it was bundled in with the popular half.
The word doing the work
The Federal Reserve Board has decided its annual stress test should surprise banks less. Two final rules, issued on Wednesday, do two separate things. They make more of the test visible in advance. And they reduce how much the capital requirement produced by the test moves from one year to the next. The first is a reform. The second is a concession, and it arrived wearing the first one’s clothes.
Start from the beginning. Large American banks must hold a cushion of shareholders’ money against losses. Part of that cushion is set by running each bank’s balance sheet through a hypothetical severe recession and seeing what is left. The worse the result, the more capital the bank must carry. The models that do this projection have historically been the Fed’s own, and not published, on the reasoning that a model you can read is a model you can arrange your portfolio around.
The Board announced in December 2024 that it would change this, proposed rules in 2025, and has now finalised them largely as proposed. Vice Chair for Supervision Michelle Bowman’s framing is that the changes keep the test “transparent, granular, and risk-sensitive”. Two of those three are new. The third is the one under pressure.
“Volatility” is the word doing the work in the Board’s own description, and it conflates two very different things. One is measurement noise: a requirement that jumps because a scenario was drawn slightly differently, or because a model quirk hit one portfolio harder this year than last. That movement is a defect and removing it is a public good. The other is genuine change in what a bank is exposed to. If a bank grows a new lending book fast, its requirement should move, sharply, in the year it grows. Smoothing does not distinguish. It flattens both.
Publishing the model has a price
The accountability case for disclosure is strong and I accept it. A rule that determines how much capital a private firm must hold is a rule, and rules in a democracy are ordinarily written down where the governed can read them and argue with them. Running the most consequential part of American bank capital policy through an unpublished model was never comfortable.
But disclosure has a cost, and the honest version of this reform states it. A published model is an optimisation target. Banks employ people whose job is to work out which assets the disclosed methodology treats gently and to shift the mix accordingly. This is not misconduct. It is arithmetic, and it is what every published capital rule has produced since the first risk weights were written down. The result is a test whose measured severity stays constant while its real severity decays, because the thing being tested is reshaped to fit.
The usual answer to that decay is discretion: the supervisor changes the scenario, changes the model, and the requirement moves. That answer is weaker once the requirement has been engineered to move less.
The risks that arrive fastest are the ones the model knows least
The timing sharpens this. The Bank of England’s Financial Policy Committee recorded this week that the risk outlook has worsened, that sovereign yields have risen and tightened conditions globally, and that rapid growth in debt issuance tied to artificial intelligence is broadening capital markets’ exposure to a single theme. Whatever one thinks of that judgement, it describes a class of exposure with almost no loss history.
A granular, historically calibrated model is good at telling you how a credit card book behaves in a downturn, because it has seen several. It is poor at telling you what happens to a concentrated book of lending against data centres, because it has seen none. Granularity is calibration on the past. When the concentration is new, granularity is not risk sensitivity; it is false precision, and a requirement damped against sudden movement will report the change last.
The Board has, to its credit, left one thing open. It is consulting on whether the test should better reflect how different banks earn fee income. That is the right instinct — a bank whose fees come from custody and a bank whose fees come from trading are not the same in a crisis — but commenters should press hard on whether historical fee stability survives the stress being modelled. Fee income that has never been tested is not stable. It is untested.
The strongest case for the other side
The case against me is real. A capital requirement that lurches by a full percentage point on model changes the bank cannot see, cannot replicate and cannot contest is not supervision, it is weather. Banks respond by holding buffers on top of buffers, which is capital not lent. Predictability genuinely buys resilience, because a firm that can plan its capital does not have to hoard against the possibility of a bad draw. And a secret model is unfalsifiable by definition; no one outside the building could ever show it was wrong.
All true. But it argues for publishing the model, not for damping the output. The Board could have done the first and, in exchange for the gaming risk that publication creates, held the average level of required capital constant or raised it. Taking the predictability and leaving the level where it was is the bargain a supervised firm would have written.
The Board should publish, with each year's results, what every bank's requirement would have been under the previous methodology — a single column of numbers showing what the smoothing costs in capital. If the answer is small, the criticism dies; if it is large, the public is entitled to know before the next downturn, not after it. And commenters on the fee income proposal, which remains open, should demand that the Board show how its proposed distinctions between business models behave for revenue lines with no recession in their history.
What would change this view
Name a number. If the Board publishes, alongside its 2027 results, a comparison showing what each large bank's stress-test capital requirement would have been under the old methodology, and the aggregate difference across the largest banks is trivial, then the smoothing is cosmetic and my objection collapses to a quibble about drafting.
Equally, if a future test produces a sharp increase in required capital for banks with concentrated exposure to AI-related lending or sovereign duration, the claim that publication and smoothing together blunt the test is wrong on the evidence. If instead no such comparison is ever published, the Board has chosen not to let anyone check.