Averaging two stress tests only works if each test is worth averaging
The Fed will set capital add-ons from a two-year average from 2028 — while admitting, in the same announcement, that the model behind fee income losses needs fixing. Smooth first, accurate later is the wrong order.
Each year the US central bank, the Federal Reserve, imagines a severe recession and works out how much money its biggest banks would lose. Those losses set how much spare capital each bank must keep. New rules let the public comment on the test, judge banks against the harsher of two market crashes, and from 2028 set the capital figure as the average of two years' results.
Averaging two years of stress test results makes a bank's capital requirement smoother, not more accurate — especially when the Fed has just admitted one of the models feeding it needs fixing. Fix the model first, then average.
Average two numbers and you get a smoother number
You do not get a better one. That is the flaw running through the changes the Federal Reserve finalised on 30 September.
Start from the beginning. Every year the Fed runs its largest banks through a hypothetical severe recession — unemployment spiking, asset prices collapsing — and calculates the losses each would suffer. Those projected losses translate into an extra slice of capital the bank must hold on top of its minimum, called the stress capital buffer. It is the one part of American bank capital rules that is tailored to the individual firm and pointed forwards rather than backwards.
The Board finalised two rules. The first opens annual scenarios, and material changes to the models behind them, to public comment, and requires banks to be tested against two global market shocks each year with the worse of the two counting. The second, effective in 2028, sets the buffer using the average of the last two annual tests. The Board says the package could halve the year-to-year volatility banks see in their stress-test capital requirements. Both rules, it notes, are largely similar to what was proposed in 2025.
The admission sitting next to the rule
Alongside the final rules, the Board requested comment on a revision to the model it uses to project noninterest income — the fee income banks earn from cards, asset management, advisory work and the rest. The stated purpose is to better capture differences in how banks generate that income. Read that as what it is: the current model does not distinguish well between business models that earn fees in very different ways. Comments are due sixty days after the proposal reaches the Federal Register.
So on the same day, the Fed locked in a rule that averages two years of test output and opened a consultation on whether one of the inputs to that output is measuring the right thing. The order is backwards. Averaging is a noise-reduction technique. It works when the errors in each reading are random and independent. It does nothing at all about a systematic bias — if a model consistently under-projects fee income losses at one type of firm and over-projects them at another, averaging two runs of the same model reproduces the bias exactly, with more confidence attached to it.
Averaging across a rule change is not averaging
There is a sharper version of this problem. Suppose the noninterest income revision is finalised sometime between now and 2028, as the timetable implies it might be. The first averaged buffer then blends a result computed on the old model with a result computed on the new one. That is not an average of two measurements. It is an average of two different measuring sticks, and no bank’s capital requirement should be half-derived from a methodology the Fed has already replaced because it was not granular enough.
The Board can avoid this. It can finalise the model first and start averaging after, or recompute the earlier year on the new basis. What it should not do is let a transparency reform — which is what the comment process on scenarios and models genuinely is — collide with a smoothing rule and produce numbers nobody can decompose.
Halving the volatility halves the signal
The Fed’s own framing is that volatility in stress-test capital requirements is a problem to be solved. Some of it is. A bank whose requirement lurches because a scenario narrative changed, rather than because the bank changed, is being asked to hold capital against the Fed’s imagination rather than its own book.
But not all movement is noise. When a bank grows a portfolio that performs badly under stress, the buffer is supposed to move. A two-year average mechanically damps that response: the year a firm’s risk profile deteriorates, only half the deterioration reaches its requirement. Bowman’s statement says the changes keep the test transparent, granular and risk-sensitive. Granularity and risk sensitivity are exactly what averaging trades away, in return for predictability. That trade may be worth making. It should be named as a trade, not presented as a free improvement.
It is fair to note the package is not one-directional. Measuring results against the worse of two global market shocks adds conservatism for trading-heavy firms, and probably a meaningful amount. The Board has not said which effect dominates for which banks, and it should.
The case for smoothing, answered
The strongest argument the other way is about behaviour rather than arithmetic. If a bank cannot forecast its requirement within a reasonable band, it holds a buffer on top of its buffer, and that management cushion is dead capital that finances nothing. Smoothing the requirement lets banks run closer to the line, which supports lending. The buffer is a floor, not a prediction, and a firm whose risk genuinely rises will see it in both years soon enough.
That argument justifies predictability. It does not justify averaging. If the problem is that banks cannot absorb a sudden increase, give them time to meet it — phase increases in over several quarters while publishing the unaveraged result in full. The bank gets its planning horizon; supervisors and creditors still see what this year’s test actually said. Averaging achieves the same cushioning by degrading the number itself, which is the one thing a capital test cannot afford to do.
The Board should sequence this properly: finalise the noninterest income model revision before the averaging rule takes effect in 2028, and commit now that the two years being averaged will be computed on the same methodology — recomputing the earlier year if necessary. Anything else blends a superseded model into a live capital requirement.
It should also publish, for every bank in every averaged cycle, both the single-year result and the two-year average side by side, with the contribution of each global market shock shown separately. If the smoothing is as modest as the Fed implies, the disclosure costs nothing. If it is not, supervisors, creditors and Congress are entitled to see how much of a firm's requirement reflects this year's risk and how much reflects last year's.
What would change this view
Show me the sequencing done properly. If the Board finalises the revised noninterest income model, applies it consistently, and either restates or recomputes the earlier of the two averaged years on that same model before the averaging rule bites in 2028 — and publishes both the single-year result and the averaged requirement for each bank — my central objection disappears.
Equally, if the first averaged cycle shows single-year and two-year-average buffers within a few tenths of a percentage point of each other across the tested firms, the smoothing is doing almost nothing and I have made too much of it. Either outcome is observable on a published date, in a published table.