The Fed’s New Global Market Shock Methodology: Consistent, Coherent, But Containing One Fatal (But Correctible) Flaw

For the first time in the 15-year history of the Federal Reserve’s annual stress test capital adequacy exercise, the Fed released extensive documentation^[1] of the models used to construct the Global Market Shock (GMS) component of the 2025 stress test and sought public comment. The GMS must be run by banks that have significant trading activity^[2] if they are subject to the supervisory stress tests under Dodd Frank.

Constructed by the Fed, the GMS is a trading book risk stress test in which financial quantities, such as equity prices or credit spreads, suddenly and violently move and markets become totally illiquid; thus, banks cannot trade or hedge their portfolios as asset prices violently move against them. The losses that a bank computes from running the GMS are incorporated into its annual capital adequacy requirement.

The new GMS documentation reveals a well-thought-out methodology that is empirically grounded and based on sophisticated econometrics. Most variables used in the GMS are determined by GARCH models,^[3] which account for volatility in financial times series data. Correlation in the GMS is handled by the so-called “copula” approach, which can capture the co-movement of extreme changes of financial variables.^[4]

But the GMS methodology has an Achilles’ heel: as a practical matter, no objective or analytically based constraints on the primary risk factor shocks that determine all other GMS shocks determined by the GARCH models. Fundamental risk factors for the five main financial markets (equities, credit, interest rates, commodities and foreign exchange) are subject to no meaningful regulatory constraints.

Thus, in practice, a small number of judgmentally determined shocks to the fundamental risk factors – called the “primary risk factor shocks” – are fed into the econometric machinery, which then produces the large number of shocks to all remaining risk factors. But there is no empirical method or theory to guide selection of the primary risk factor shocks. They could be set to assume market movements of unprecedented individual and collective severity, without explanation and therefore without any meaningful ability for the public to comment.

In this research note, we review how the GMS methodology works. Then we will implement it for a credit example, showing how arbitrarily small changes in assumptions on primary variable shocks produce large changes in the GMS credit spread shocks. We choose a credit example since credit shocks are disproportionately important for determining losses to banks. We conclude with some discussion about how the GMS methodology could be improved to ensure greater objectivity and robustness.

How The GMS Methodology Works

The GMS is a stress test with a very large number of shocks that must be specified. A shock is a sudden change in a financial price. For example, a 200-basis-point shock for the BBB credit spread requires a bank to add 200 basis points to the current BBB credit spread and then compute the profits or losses on all the BBB credit instruments in its portfolio. Alternatively, an equity shock might be a sudden 20 percent decline in the S&P 500 index, requiring a bank to compute resulting gains or losses.

To keep the process tractable, the GMS methodology proceeds in three stages. First, Fed staff defines judgmentally a small number of primary variable shocks that are representative of the main financial markets.^[5] For credit markets, an example of a primary shock is the Bbb-Aaa spread. Second, a simple regression models to infer “secondary” variable shocks from the primary variable shocks. Secondary shocks are shocks of variables related by a regression model to the primary shocks. An example of the secondary shock in the credit markets is a shock to the BBB spread. Lastly, the primary and/or secondary shocks are fed into GARCH^[6] models to construct the large number of remaining shocks. Figure 1 depicts the process:

Figure 1

Stage 1: The Primary Shocks

The Fed chooses the primary shock variables to broadly cover the five main financial markets: equities, credit, interest rates, commodities and foreign exchange. The selection of the primary shock variables might vary from year to year. For the 2025 stress test, the Fed chose the S&P 500 index, the Moody’s Baa – Aaa spread, the level and slope of U.S. Treasury rates, assorted commodity and energy indices and the USD to euro exchange rate.

Once the primary variables are chosen, the Fed must specify the shocks to these variables. According to the GMS documentation, the Fed writes a scenario narrative to guide the selection of the shocks. The Fed then classifies the severity of the shocks qualitatively as being: 

• Mild
• Moderate
• Large
• Severe
• Unprecedented

Each category specifies a range of shocks that may be chosen. For this purpose, the Fed defines two liquidity horizons, one month and three months. The liquidity horizon is the period over which the Fed assumes that the market has no liquidity in each asset class – meaning that no trades or hedges are possible.

Credit, for example, has a 3-month liquidity horizon, which means that under the GMS, credit spreads are assumed to move over a 3-month period while banks are powerless to sell or hedge losses to credit instruments they hold. Given that the primary risk factor for credit is Baa-Aaa and the Fed judgmentally defines credit shocks to be in the “Large” category for a particular stress test, then the shock could be chosen anywhere between the 95^th and 99th percentile 3-month move that was observed historically. If the primary credit risk factor shock were classified as “Severe,” the shock would be anywhere between the 99^th percentile 3-month historical move and the maximum 3-month historical move. An “Unprecedented” credit shock is defined to be greater than the maximum 3-month move observed historically. Once the Baa-Aaa risk factor is established, the secondary credit shocks are set at the same/ similar level of severity, and the remaining credit shocks derived from them. 

The methodology works the same for other asset classes, although the liquidity horizon may be different and the direction may reverse from an increase to a decline. For example, equities have a 1-month liquidity horizon. A “Large” shock to the S&P 500 would be a 1-month percent move that is between the 1^st and 5^th percentile of the historical record, selecting large downward 1-month equity moves.

At present, the first stage is a black box. The Fed does not reveal the risk category (Mild to Unprecedented) in which it has placed each of the primary shocks; what percentiles it chose within the chosen risk category; or over what historical time frame the primary shock sizes were computed.  It also does not disclose why any of those choices were made or how they interrelate.

Note that an assumed three-month illiquidity period is already a draconian assumption without precedent for many types of credit. Assuming a 95-99^th percentile for any of the five primary risk factors would be a conservative assumption but note that staff is free to assume that same shock for all five primary risk factors, which has no precedent, creating a massive capital requirement. Even then, staff could abandon historical precedent altogether and impose Unprecedented shocks for the primary risk factors.

While future scenarios are expected to be published for public comment as a potential check against such excess, such comment is unlikely to be meaningful, as there is no narrative to assess; no standard against which to measure the scenario, and the severity of the shocks for each risk factor is not disclosed.

Stage 2: The Secondary Shocks

Once the primary shocks are chosen, the Fed estimates a series of regressions to derive the secondary shocks from the primary shocks, in most cases employing quantile regressions. For example, once the Fed has defined the primary shock to the Moody’s Baa-Aaa spread, it would substitute the primary shock into an estimated quantile regression to derive the secondary shock to the BBB bond spread. Each primary variable shock has at least one secondary shocks attached to it, each of which is determined by a regression model.  The Federal Reserve justifies the use of quantile regressions on page 25 of the GMS model documentation, saying “In market risk, extreme shocks tend to happen simultaneously during financial crises. This behavior is captured by the quantile regression model because it expresses the conditional quantiles of secondary risk factors as a function of primary risk factors.” Similarly, on page 66, it is asserted that “…the quantile regression model described in Section C.ii.1.a are designed to capture tail outcomes of the dependent variable.”

These justifications are conceptually incorrect. A quantile regression does not capture correlation between extreme shocks or tail dependency. A quantile regression merely measures the effect of any change in the independent variable, whether small or large, on the kth percentile of the dependent variable.^[7] The quantile regression answers the question: what is the effect of a marginal change in one financial quantity on observations of another financial quantity that are in the τ^th percentile? The quantile regression does not describe the joint probability of a simultaneously large shock to both variables. It does not imply that a large shock to one variable must necessarily follow a large shock to a different variable.

Stage 3: The GARCH t-Copula Methodology

Once the primary and secondary risk shocks have been specified, the Fed then turns to the last stage in which most of the GMS shocks are determined. The GARCH t-copula econometric methodology is used to ensure empirical consistency and coherence of the large number of related shocks with the primary and secondary shocks.

As an example of how this would work for credit spreads, the Fed would pick a primary shock for the Moodys Baa-Aaa spread. It would then derive the secondary shock for the BBB spread using a quantile regression. Finally, it would calculate the implied shocks to the bond index AA-AAA, A-AA, BBB-A, BB-BBB, B-BB, and CCC-B spreads by inserting the BBB bond index shock into a conditional GARCH t-copula econometric model estimated for each spread.

We implemented the Fed’s GMS methodology for the credit spread example presented in Tables C3, C6, and C7 in the technical document. To obtain the monthly Moody’s Bbb-Aaa spread, we subtracted the Moody’s Aaa yield from the Bbb yield from 1997 to the present, both series taken from FRED. We also obtained the monthly BBB ICE option-adjusted spread over the same period from FRED. We then estimated the quantile regression between BBB and Bbb-Aaa at the 90^th percentile as in Table C3.  Next, using weekly data from 2005 to present from FRED, we constructed credit spreads and then estimated the GARCH t-copula model presented in Table C6 and also the Kendall’s tau correlation matrix in Table C7 from Equation C14. We obtained estimates similar but not identical to those presented in those tables, likely because the data sets are somewhat different. We then developed the conditional simulation models in Equations C7, C9, C11, C12 and C13 from the technical document.

Armed with a fully estimated model, we can then study the sensitivity of the model to assumptions on the primary variable shock, the Bbb-Aaa spread. We make an assumption about shock classification, “Moderate,” Large,” or “Severe” and then choose a shock based on a permitted confidence level in the classification. For example, if the shock is “Large,” then we can choose any percentile between 95^th and 99^th percentile. We calculate the primary variable shock by choosing an empirical confidence level on weekly data over a 3-month horizon using two time periods, 1997 – present, and 2005 to the present. Once we have the primary shock, we substitute it into the estimated quantile regression to find the secondary BBB shock. Finally, we substitute the BBB shock into the conditional GARCH t-copula model and perform 10,000 simulations, taking the average shock results for each of the credit categories.

Table 1

Table 1 reports the results of a Moderate shock for the 85^th, 90^th, and 95^th confidence level and for two time periods of the data. As is apparent, even if a shock is Moderate, the shock size can vary significantly, depending on the confidence level chosen. However, the effect of using different time periods is small.

Table 2

Table 2 shows that very small changes in percentiles for Large shocks can significantly change the GMS. Moreover, at the top of the Large shock band, 99^th percentile, the historical period used to calculate the shock matters significantly. At the 99^th percentile and above, there are relatively few shocks. Adding more data will move the 99^th percentile shock to a lower level, because there are more points in the upper tail of the historical distribution.

Table 3 shows the results for Severe shocks and compares to the 2025 GMS for Bond indices and single name CDS. The Fed implements the methodology separately for bond and CDS spread shocks.

Table 3

Since we used bond spreads to compute the GMS spread shocks, and bond shocks tend to be greater than CDS shocks, a comparison of the example results and CDS shocks in Table 3 suggests that the Fed is classifying the 2025 GMS CDS spread shocks with respect to a primary variable shock somewhere in the “Severe” category or greater. It is unclear what classification the Fed is using for its bond GMS shocks. In the “Severe” category, extremely small changes in assumptions about percentiles can have a very large effect on the resulting GMS. For the lower percentiles in the Severe category, the time period used to calculate primary risk factor shock matters as well. At the upper percentiles, such as 99.7, the time period is not as important, since at such a high confidence level there are very few data points left in the upper tail.

Tables 1, 2, and 3 show that the Fed’s GMS methodology allows the Fed scenario designers to set the GMS severity, and thus the capital required for trading, at whatever level they want. There are no objective rules, analysis, or empirical justification, and no constraints on picking any primary risk classification, and within those risk categories, any percentile and historical time frame. Within the risk category that the Fed seems to have chosen for the 2025 credit portion of the GMS, “Severe,” or greater, extremely small variations in percentiles or the historical time frame to calculate the shocks can have a very large effect on the GMS results.

The Fed has given itself even more flexibility to make the GMS whatever it wants with its “Unprecedented” category. If the GMS is not considered to be severe enough, the scenario designer can arbitrarily pick any primary risk factor shock, set it at a level outside of the historical data, and feed it into the methodology. Table 4 shows an example in which the Bbb-Aaa shock was chosen such that the BBB shock that emerged from the quantile regression was 450 basis points. Plugging the BBB shock into the GARCH t-copula produced a very severe GMS.

Table 4

Improvements to the GMS Methodology

The weakest spot in the GMS is the selection of the primary risk factor shocks. In this section, we suggest some potential improvements, with most being to the first stage of the process.

Governance and Transparency

The Fed should publish its reasoning around the selection of the primary shocks, including the scenario narrative, before each stress test exercise and explain how its choices are related to the narrative, subject to public comment.

Reduce the Degrees of Freedom in the Methodology

The choices in the Large and Severe categories for the Primary Risk Factors should be subject to some objective constraints, since scenario designers cannot really distinguish between small differences in percentiles that can have such a large effect on the results. The Large shock should be at the 95^th percentile rather than a range, and a Severe shock should be set at the 99^th percentile. These are standard risk management percentiles. Given that the methodology already assumes that market liquidity has been eliminated for one or three months, which is a highly unrealistic and draconian assumption, it is difficult to justify setting confidence levels about standard, very conservative levels.

It would be important to keep the “Unprecedented” risk category to maintain scenario flexibility in unusual market circumstances. But there should be publicly disclosed rules and a methodology around its use that would limit its application to rare market events, such as the Covid outbreak. Any use of the “Unprecedented” category should be publicly reported and documented in the scenario release.

These guardrails around GMS severity take on increased importance in light of the agencies’ approach to addressing potential areas of overcapitalization between the supervisory stress test (the GMS) and regulatory capital requirements under the proposed Basel-based Enhanced Risk-Based Approach (Basel III).  Rather than eliminating the capitalization for market risk under either the stress test or the Basel III requirements, instead the agencies take an overall calibration approach, noting that “[t]he increase in market risk-weighted assets as a result of the Basel III proposal would be more than offset by the proposed changes to the global market shock component of the Board’s stress test.”^[8] The Federal Reserve’s emphasis on overall calibration of market risk requirements to eliminate inappropriate overcapitalization across the capital framework reinforces the need for guardrails on GMS severity.

Rationalize the Liquidity Horizons

As discussed in the recently released model documentation, the Fed has historically assumed long liquidity horizons that varied by asset class, typically six months or longer for credit shocks for example. In 2025 GMS exercise, the Fed greatly improved its methodology by standardizing liquidity horizons of one month or three months, inspired by the Fundamental Review of the Trading Book (FRTB) liquidity horizons. Nonetheless, the GMS horizons are still arbitrary and overly severe, since the FRTB horizons they are based on were never properly justified.

The horizons could be rationalized by developing a clear, realistic theory of the nature of the assumed market illiquidity in an asset class. Hopper (2025) referenced earlier suggests a limited liquidity methodology in which a small amount of a trade can be sold or hedged per day. That methodology is fully consistent with the GARCH models the Fed already employs. It is also very flexible, allowing for different specifications of the dynamics of daily limited liquidity.

Exploit the Flexibility on Correlation in the Model, But Be Careful to Treat it Reasonably

The GMS methodology grants scenario designers infinite flexibility to set the severity of the GMS, and thus the level of capital, at any level. The severity can be changed willy nilly, while the t-copula maintains tight, but conservative control of the correlations of extreme moves. However, a robust methodology should have the ability to break empirical correlations in a stress test. The Fed’s methodology is well-suited to that task. The Fed has many levers it can use to break correlations. For example, it can change the estimated t-copula tail dependency to turn up or down tail dependence of extreme moves. It can also change the estimated correlation matrix. And it can choose which variables to keep in or out of a specific t-copula, thus affecting their correlation.

In doing so, however, the Fed must be careful not to unjustifiably break historical correlations. As a simple example, interest rates in two countries are related to forward exchange rates through covered interest parity. Covered interest parity is an arbitrage-free condition on the relationship between interest rates and forward exchange rates. It might have small violations, but if the relationship between interest rates and forward exchange rates gets too out whack, it would imply large arbitrage profits are possible in the scenario.

A real-life example of this problem can be seen in the large discrepancy between the bond and CDS credit shocks in the 2025 GMS. CDS credit spreads are derived from bond spreads, so they should not dislocate to the extent they do in the 2025 GMS. We do not know why the bond and CDS shocks are inconsistent, since the Fed has not documented how the shocks were set. One explanation for the discrepancy may be that the Fed included bond shocks in one t-copula GARCH model and CDS shocks in a different model. That would effectively break the correlations between them. But without documentation, we can only speculate.

The Fed should disclose its thinking on correlations, including which variables are assumed to be correlated within the same t-copula. If it decides to break empirical correlations in a GMS exercise, it should document where it has done that and the reasons for it. 

Consistency with the FRTB

Once Basel is finalized, the FRTB will introduce a market risk model that stresses the risk that the GMS emphasizes now – periods of long market illiquidity. The FRTB is based on an explicit or standardized market risk model that assumes extreme market illiquidity. The purpose of a stress test is not to repeat the risks covered by the FRTB market risk model but to stress risks that are not or cannot be captured by the FRTB model. Accordingly, once FRTB is in place, the Fed should move most shocks that are currently in the Large or Severe categories to the Moderate or Mild categories, since the FRTB is already assuming high confidence level shocks over long periods of illiquidity. The GMS stress test scenario design should then concentrate more on probing for different risks that banks might face. The ability to robustly, but reasonably, break correlations in the GMS stress test will be a key capability.

Conclusion

The GMS methodology has a sophisticated and well-justified econometric engine under the hood, but it is ultimately nothing more than a series of dials that can be adjusted to set the GMS shocks, and thus the level of capital, to any level desired. Of course, a scenario design methodology requires flexibility to be effective, but infinite severity flexibility defeats the purpose of ensuring that the GMS is plausible and objective. In this research note, we made a series of recommendations that can improve the methodology without requiring fundamental changes.

In short, the Fed needs to improve the justification, transparency and governance of the small number of choices the Fed makes that essentially determine the severity of the GMS. Once Basel is finalized, it will also be important to adjust the calibration of the GMS to stress different rather than the same risks the FRTB emphasizes.

The Fed should be commended for developing a very nice stress testing methodology that offers some important innovations. But it is very important to make some simple changes so that the GMS does not degenerate into a random capital generator, moving up and down from year to year for arbitrary reasons that no one understands. 

---

[1] Board of Governors of the Federal Reserve System, “Supervisory Stress Test, Global Market Shock Component,” October 2025, available at https://www.federalreserve.gov/supervisionreg/files/gms-model.pdf

[2] “Significant trading activity” is defined as (1) aggregate trading assets and liabilities of 50 billion dollars or more; or (2) aggregate trading assets and liabilities equal to 10 percent or more of its total consolidated assets.  12 CFR 252.53(b)(2).

[3] Generalized Autogregressive Conditional Heteroskedasity.  They are as complicated as the name suggests.

[4] The use of GARCH models and copulas in the GMS were advocated by Hopper, Greg, “How to Make the Global Market Shock Coherent,” Bank Policy Institute, May 2025 available at https://bpi.com/how-to-make-the-global-market-shock-coherent/ and Hopper, Greg, “Rationalizing the Global Market Shock,” Bank Policy Institute October 2023, available at https://bpi.com/rationalizing-the-global-market-shock/. See these notes for complementary discussion.

[5] Note the shocks historically have been set by the Fed staff with review by the Vice Chair of Supervision, but the Board has never reviewed or approved the scenarios or any of their components.

[6] For a discussion on GARCH models, see Hopper, G, “Rationalizing the Global Market Shock,” Bank Policy Institute, 2023, available at https://bpi.com/rationalizing-the-global-market-shock/

[7] The quantile regression may be written as Q_τ(Y|X = x) = α_τ + β_τx

[8] Federal Reserve Board, Office of the Comptroller of the Currency, Federal Deposit Insurance Corporation, Regulatory Capital Rule: Category I and II Banking Organizations, Banking Organizations with Significant Trading Activity, and Optional Adoption for Other Banking Organizations, 91 Fed, Reg. 14952, 15104.  Indeed, the agencies further state in the Basel III proposal that “[o]verall risk-weighted assets for trading-related activities of Category I and II banking organizations would rise 30.9 percent (see table VII.5). Despite this significant increase in risk-weighted assets, when considering the other proposals, in particular the proposed changes to the global market shock, the combined impact on capital requirements would be a moderate 9.2 percent.”  Id. at 15138.