There are basically no pictures of him not smiling.

Photographer: Jemal Countess/Getty Images

Goldman Sachs Is Cutting It Close on the Stress Test

Matt Levine is a Bloomberg View columnist. He was an editor of Dealbreaker, an investment banker at Goldman Sachs, a mergers and acquisitions lawyer at Wachtell, Lipton, Rosen & Katz and a clerk for the U.S. Court of Appeals for the Third Circuit.
Read More.
a | A

Will Goldman Sachs fail today's stress test? Well, what would that mean? Last week's stress test results measured whether banks were adequately capitalized. At this point, six years into the bull market, it would be a huge embarrassment if any weren't. And in fact they all were. 

This afternoon's results, on the other hand, determine whether banks can return capital to shareholders through new stock buybacks or dividend increases. "Failing" this test means only that you can't return any more capital. "Passing" this test means that you can return the amount of capital that you asked to return. And you had two shots to ask for an amount of capital return: one before you saw last week's results and one after.

This means that the test is graded on an oddly generous curve. If you initially asked to return $5 billion of capital, and last week's results show that you can return $4 billion, then you revise your ask to, like, $3 billion. (You don't want to look greedy, after already having gotten it wrong.) Goldman did something like this last year, revising its request after initial negative feedback. On the other hand, if you initially asked to return $5 billion of capital, and last week's results show that you can return $20 billion, then you stand pat. (It is not allowed, or at least not done, to increase your ask.) And if you initially asked to return $5 billion of capital, and last week's results show that you can only return $700 million -- as seems to have happened to Goldman -- then you revise your ask to, I mean, I think humility calls for a revised ask of like, "We'll buy back one share." Or like, "Our capital return plan is, we're good for now." But the point is, you don't have to fail. You can just ask for less. If you ask for too much, and the Fed gives you nothing, then you "failed." If you ask for nothing, and the Fed gives you nothing, then you've "passed." And since last week, you know how much the Fed will give you.

It would take some perversity for Goldman to actually fail, though of course one can never rule out perversity.

In addition to perversity, the other possible confounding factor is a "qualitative" failure, which is when the Fed rejects your capital plan not because it leaves you with too little capital, but because it thinks you're generally inept at capital planning, or doesn't like the smug look on your face during this stress test. I guess that could happen to anyone -- it happened to Citi last year, and Goldman has occasionally been spotted with a smug look on its face -- though the speculation about qualitative failures this year seems to focus mostly on U.S. units of foreign banks.

Of course "passing" by getting a tiny or no buyback approved feels like a failure. But look on the bright side: There is a sense -- a very unsatisfying sense, sure -- in which Goldman has already aced the stress test, by having just barely enough capital to meet regulatory requirements. It already returned the most possible capital! All the other banks that are way over the minimums, and now asking to return more capital to shareholders, are doing it wrong. It's like a tax refund: Their shareholders have been making an interest-free loan to the capital regulators.

I guess no one believes that? Part of the problem is that the stress test isn't just a measure of whether Goldman is adequately capitalized (it is), or of whether it can return capital (meh?), but of how closely its thinking matches the Fed's. John Carney has a great column about how Morgan Stanley and, especially, Goldman have views of their stressed capital that are very different from the Fed's view. This looks sort of bad: The Fed grades the test, and if Goldman's models diverge wildly from the Fed's, then that might mean that, in some regulatory sense, Goldman's models are "wrong."

Carney doesn't buy that:

In fact, Fed officials have explicitly stated that they don’t want banks to construct portfolios optimized for stress testing. Nor do they want banks to create models meant to mimic the Fed’s, which one Fed official compared with schools “teaching to the test.” Instead, banks are expected to make their initial capital requests based on models that reflect their own well-grounded expectations and assumptions.

He attributes the divergence here mostly to differing models of trading performance:

Long term, though, this doesn’t signal the kind of failing that should really alarm investors. Rather, it reflects the continued importance of trading to Goldman’s and Morgan’s businesses, even despite the latter’s attempts to diversify.

In the tests, even small variations in financial models related to trading performance can produce disparate results. Given that, both firms are likely to continue seeing wider variations with Fed views during annual stress tests.

Here's an approximate model I put together for how Goldman and the Fed get from Goldman's current 19.8 percent total risk-based capital to a nadir of (for Goldman) 13.0 percent or (for the Fed) 8.1 percent, just barely above the 8 percent requirement:

One thing to notice here is that Goldman and the Fed are actually pretty close to each other on trading and counterparty losses: The Fed sees $17 billion of losses; Goldman is higher, at $19.1 billion, but the difference doesn't have a huge effect on the outcome. In fact, Fed and bank models on trading losses generally tie out pretty well; of the usual six big banks, model differences over trading losses had the biggest effect on capital at Goldman, and even that wasn't huge:

Source: Banks, Federal Reserve, my calculations

The much bigger effects come from how the banks and the Fed model revenue and risk-weighted assets differently:

I think it's worth trying to draw a very hazy distinction between models of losses on things you currently own (trading assets, loans), on the one hand, and models of future revenue or asset growth, on the other. The first category is in some sense uncontrollable. Like, this is a stress test. It's meant to test what happens when you have losses because the economy goes bad, market assets drop in value, and your counterparties default. You can't answer, "Oh well all our loans would do well, even in a crisis, because we only lend to the best people." The whole idea of the test is that your loans, trading assets, etc., are blowing up, in a way that the Fed has pretty explicitly quantified in the stress-test scenario.  So you model that, and the Fed models that, and you use historical data, and it's all fairly mathematical and you should come to about the same answer.

But modeling revenue and asset growth is importantly different. Here's a sample of how the Fed models pre-provision net revenue, from page 61 of last week's stress test release:

PPNR is forecast with a mix of structural models using granular data on individual positions: autoregressive models that relate the components of a BHC’s revenues and non-credit-related expenses, expressed as a share of relevant asset or liability balances, to BHC characteristics and to macroeconomic variables; and simple models based on recent firm-level performance. ...

Trading revenues are volatile because they include both changes in the market value of trading assets and fees from market-making activities. Forecasts of PPNR from trading activities at the six BHCs subject to the global market shock are modeled in the aggregate and then allocated to each BHC based on a measure of the BHC’s market share. 

Here's how Goldman models revenue, from page 6 of its release:

When projecting these business segment revenues, we utilize multiple approaches, including models based on regression analyses, management judgment and projecting the impact of re-pricing inventory due to the projected changes in asset values under the Federal Reserve Board’s severely adverse scenario. We also incorporate the impact of broader industry performance during historical stressed periods to help guide management judgment regarding our future performance in the assumed stressed operating environment. The projected revenues under the Federal Reserve Board’s severely adverse scenario are an aggregation of projected revenues for each of these business segments.

The Fed's approach is largely statistical, based on historical data: It models revenue sort of the same way it models trading losses. The banks' approach is managerial, based on historical data but also on being a bank and planning for things. Revenue, for Goldman, is not the passive result of a statistical process; it is a thing that is managed and planned for. 

Similarly for assets. The Fed (page 62):

The BHC balance sheet is projected based on a model that relates industrywide loan and non-loan asset growth to each other and to broader economic variables including a proxy for loan supply. The model allows for both long-run relationships between the industry aggregates and macroeconomic variables, as well as short-term dynamics that cause deviations from these relationships.

Goldman (page 7): 

Projected RWAs reflect the impact of the macroeconomic environment; for example, changes in volatility and credit spreads are incorporated into our calculation of projected RWAs. Additionally, projected RWAs and capital deductions are also impacted by the projected size and composition of our balance sheet over the planning horizon.

Goldman considers its balance sheet managerially: It can actually decide to buy and sell assets. The Fed, again, takes a purely statistical approach, treating the banks as robotically growing their balance sheets in line with current market share.

Now of course it's entirely possible that the Fed's approach is more accurate: Managers might well overestimate their ability to manage to a good result. Certainly as a capital regulatory tool, using neutral statistical models seems more reliable than relying on management judgment. On the other hand, it does seem a little hard to fault the banks' managers for thinking that they can have some effect on how the bank will perform even in a stressed situation. Isn't that kind of what they're there for?

  1. Disclosure, I still used to work at Goldman!

  2. This confuses people, because it seems like the stress test should create some failures, but that is not its purpose. Its purpose is to increase bank capital. Here is a good way of putting it:

    Stress tests, rather than a non-credible show of invincibility, are just another set of capital requirements.

  3. No, I kid. There are timing issues: Goldman barely scrapes over the capital minimums at its lowest point, but it's comfortably over the minimum at the end of the stress-test period. So analysts seem to think that Goldman can structure a capital-return plan to do most of its buyback toward the end of the period, so that it never dips below the requirements, but buys back more than, like, nothing.

  4. Good lord don't think too hard about this analogy. Anat Admati, I am kidding!

  5. Okay so. The raw data comes from Goldman's and the Fed's releases of Dodd-Frank Act Stress Test results from last week. I've put in dividend data, assuming nine quarters of dividends at 60 cents a share on 435.6 million shares. The categories are:

    • "Revenue" corresponds to the Fed's categories of pre-provision net revenue and other revenue.
    • "Loan losses" is provisions.
    • "Trading losses" is trading and counterparty losses.
    • "Other losses" is "realized losses/gains on securities (AFS/HTM)," plus "other losses/gains," plus "other comprehensive income" to the extent it flows through capital.
    • "Dividends" is my calculation.
    • "Change in assets" is the effect on capital ratios -- after accounting for the above revenue and losses -- of the increase (or decrease) in risk-weighted assets from 2014 to 2016 under the general approach.
    • "Basel III effect" is the effect on capital ratios -- after all of the above -- of the increase in risk-weighted assets from the 2016 general approach to the 2016 standardized approach. See pages 15-16 of the Fed release for more on this.
    • "Other" is just a plug. It could be operating expenses, tax effects, rounding error, capital actions other than dividends that are incorporated in the DFAST model, whatever my dumb math isn't picking up.
    • "Final" is the total stressed risk-based capital as of the end of the stress-test period, and "Minimum" is the lowest stressed risk-based capital during the stress-test period.
    • "Timing effect" is just the difference between those two numbers: The fact that this effect is big in the Fed's numbers means, roughly, that the Fed thinks that things will be relatively worse (more losses, more assets, etc.) early on in the period than Goldman does.

    All very unscientific, not investment advice, etc.

  6. Don't take this any more seriously than the other chart. Note that the effect on capital is percentage points of final total risk-based capital, using the Fed's final standardized-approach assets as the denominator.

  7. Continuing disclaimers! Very rough math! Note that the "effect on stressed capital" in the asset model is based on dividing my estimate of the total risk-based capital (end of period) as calculated by the Fed (not the banks) into the differing asset numbers.

  8. See pages 5-8 of last week's release for details on what the "severely adverse scenario" looks like.

This column does not necessarily reflect the opinion of Bloomberg View's editorial board or Bloomberg LP, its owners and investors.

To contact the author on this story:
Matt Levine at

To contact the editor on this story:
Zara Kessler at