Whales don’t live in the tail: A Cover-1 in time saves nine

9th July 2026 

Key takeaways

• A gap currently exists at the intersection of Counterparty Credit Risk (CCR) and Prudential Valuation (PruVal): the endogenous liquidation cost of large portfolios is currently not captured by either framework, thus leaving institutions at risk of large losses upon the default of a counterparty. Drawing on established frameworks from central counterparty risk management, I propose a novel “Cover-1” capital charge¹ and a limit, similar to the large exposure framework, to price and internalise contingent liquidation risks.
• The technology to use stress testing for vulnerability identification at scale already exists today and should be leveraged more, both at the micro and macro scale; such exploratory exercises should however not drive capital requirements. Rather, they should feed into fire drills, preparedness tests, and board-level discussions.
• Beyond exploratory analysis, stress testing can and should continue to be used to set microprudential capital requirements. The most important enhancements in this realm would be to: (i) increase the number of scenarios to at least two or three, and (ii) critically review internal models, especially in jurisdictions that have not experienced significant GDP contractions or loan losses in recent decades; models calibrated to benign times may severely underestimate the impact of a severe macroeconomic shock. 

Introduction 

I am honoured to share some thoughts in this Forum on how stress testing can be improved.

The previous contributions can broadly be grouped into two camps. Til Schuermann and Fernando de la Mora focus on improving the existing microprudential stress tests, in particular the European exercise: they would preserve the capital-setting and “certifying” nature of these tests, while strengthening both process and outcomes. Meanwhile, David Aikman and Laura Valderrama argue for less certification in favour of more «exploration»: stress tests – and reverse stress tests in particular – should be used to identify (new) vulnerabilities, where endogenous effects such as solvency–liquidity spirals and the role of trust should receive more prominence.

I propose the following dividing line between the previous contributions: A bank’s own footprint in a market, i.e. the endogenous price impact of liquidating its own positions, is measurable from the bank’s own data, attributable to its own choices, and therefore capitalisable (net of margin). I suggest a Cover-1 charge to that end, which to the best of my knowledge has not yet been proposed in this form. By contrast, system-wide endogenous risk – the amplification that propagates through other institutions’ balance sheets, funding decisions and reactions – is neither measurable with the precision that capital-setting requires nor attributable to any single firm. Therefore, the second belongs in exploration via vulnerability analyses, fire drills, and macroprudential surveillance rather than into capital. In short: price what is yours, explore what is everyone’s.

A well-designed Cover-1 charge to help pull whales out of the tail

Chance favours the prepared mind – Louis Pasteur

In 2012, the London Whale provided a prime illustration of endogenous liquidation costs and their potential to become of first-order importance. Cont and Wagalath (2016) show that against a USD 10.2bn liquidation-adjusted VaR, the USD 6.2bn loss no longer looks like a black swan event, which it might have done at first sight against a roughly USD 0.5bn “standard” VaR. The Whale was a current, first-party concentration: precisely the case that PruVal’s Concentrated Positions Additional Valuation Adjustment (AVA) is designed to capture today.

Valderrama highlights the Archegos losses at Credit Suisse as an example underpinning the importance of capturing wider market liquidity risks, including endogenous liquidation costs. I agree, but would press the point: as of today, neither the counterparty credit risk, nor the PruVal frameworks catch this contingent concentration risk.

Indeed, the issue is not caught by CCR: under both the standardised approach to counterparty credit risk (SA-CCR) and the internal model method (IMM), endogenous liquidation costs play no role. Price dynamics are calibrated to the past and treated as exogenous, and the margin period of risk (MPOR) does not depend on the size of the position relative to the market.² 

PruVal does not close this gap either. While PruVal already introduces the main principle, i.e. it scales prudent exit horizons by a position’s size relative to average daily volume (ADV), it only does so for positions that are concentrated today – not for those that might become so tomorrow. In the Archegos case, the AVA would have been negligible: Credit Suisse hedged the Archegos total return swaps by holding the corresponding equities, leaving the net position roughly flat. The concentration became apparent only once Archegos defaulted and the bank was left to liquidate a large outright equity position.

One may object that Archegos was also concentrated across prime brokers, and that each bank could, obviously, only see its own slice and thus question whether the above proposal solves the issue. A charge computed on each bank’s own book would, however, at least partially, have resolved the issue as Credit Suisse’s own positions alone amounted to multiple days of ADV in the key names. Hence, a charge on the own-book component would have bitten regardless. The residual problem – visibility of a client’s aggregate footprint across dealers – is one of disclosure and supervisory data sharing (the post-Archegos proposals on security-based swap position reporting point in this direction); no capital charge, however designed, can solve it. The Cover-1 charge addresses the part of the risk that lies within each bank’s control and information set.

How to fix it?

The charge. Central counterparties must, under their Cover-1 or Cover-2 requirements, be able to withstand the default of their largest one or two clearing members. Borrowing this terminology, I propose a Cover-1 charge for contingent market risk at the intersection of CCR and PruVal. The charge equals the liquidation-adjusted VaR (LVaR), at a stated confidence level, of exiting the positions the bank would be left holding upon the default of the counterparty for which this quantity is largest – net of the initial margin received from that counterparty (variation margin being already reflected in current marks). If margin is comparatively thin – as it was at Archegos – the residual charge is correspondingly larger. Where margin is posted in securities rather than cash, its value should be assessed under the same liquidation-adjusted lens as the exposure itself, correcting for any potential wrong-way risk.

Why margin structurally undershoots. Initial margin is sized for a margin period of risk measured in days; a position of five times ADV implies a liquidation horizon measured in months. The Cover-1 charge prices exactly this mismatch – the part of the risk that current margining practice, by construction, do not reach.

The incentive. The charge gives the bank a choice: hold the capital, or pass the cost on to the counterparty through higher margin and prevent the build-up in the first place. This is more than a side benefit. Archegos margins were thin because prime brokers competed them away; bilateral margin discipline failed for precisely that reason, and a Pillar 2 response would inherit the same unevenness across firms and supervisors. A uniform charge is competitively neutral and self-enforcing. It also has precedent: CCPs such as LCH, Eurex and CME already embed concentration add-ons in initial margin that price liquidation cost against market depth, and the Basel Committee’s 2024 guidelines for counterparty credit risk management state the principle qualitatively.

Measurement. To avoid the notoriously difficult task of estimating the price impact of a distressed liquidation directly, one can follow the principles underpinning PruVal: assume a conservative participation rate relative to ADV, derive the implied liquidation horizon, and apply a VaR or potential future exposure (PFE) methodology over that horizon. A stylised example illustrates the order of magnitude: a position of five times ADV, exited at a ten per cent participation rate, takes roughly fifty trading days to unwind – leading to a doubling or tripling of a static 10-day initial margin requirement.

Computation and procyclicality. The exercise need not create material overhead. Price impact is sharply non-linear in size relative to ADV. This means the term is de minimis for all but the largest counterparties – the whale logic, and the reason a Cover-1 (or small Cover-N) design suffices rather than a charge across all counterparties. Ranking counterparties by, say, the largest single-name position-to-ADV ratio within each portfolio reveals where the contingent concentrations sit. To avoid procyclicality issues of ADV collapses during times of stress a “through-the-cycle ADV” including floors could be used.

Large contingent exposure limits. In addition to the proposed Cover-1 charge that builds capital to absorb losses, a large contingent exposure limit would provide an additional safety net to prevent the build-up of excessively large positions against any single counterparty. This addresses both the issue that a “Cover-N” charge is economically unsustainable and leans against cognitive biases. Confirmation bias, among others, can lead to rationalisation of why the default of each individual counterparty is very remote, and miss the fact that it becomes increasingly likely that at least one counterparty will eventually default across a large portfolio. 

Whether the timing for introducing such new charges is right is beyond the scope of this contribution, but one could start by attempting to quantify the risk, as it would not only capture interbank exposures, but also correct the pricing of the leverage provided to the NBFI sector. 

Vulnerability analyses and reverse stress testing

Dessine-moi un mouton. – The Little Prince

Next, I’ll discuss why I believe that despite having just proposed the capitalization of own endogenous risk, the extension to the capitalization of endogenous systemic risk is a bad idea. 

Despite many advances, the vulnerability-identifying potential of stress tests appears underexplored to date. The root cause, it seems to me, is the urge to chase illusory precision at the expense of the main point. Back in 2011, Haldane and May elegantly drew lessons from ecology for finance; in this Forum’s own backyard, Danielsson and Shin coined the term “endogenous risk” as early as 2003. Yet stress testing has not fully absorbed what students of complex systems have long known. Indeed, run ceteris paribus, stress tests regularly hold fixed the very things that often decide survival of a firm: trust, access to funding, (long-term) profitability, and the reactions of counterparties and peers to adverse news. Leaving those aspects away is not a simplification. No, it actually assumes away the key mechanisms of contagion³. A population model without feedback predicts exponential growth ending in Malthusian catastrophe; but we know that is not how populations behave. The feedback-blind model is thus integrating the wrong equation at an arbitrary precision. With respect to stress testing I strongly feel that John Tukey’s point still stands that it is preferable to have an approximate answer to the right question than to have an exact answer to the wrong question. While important advances have been made in recent years in the modelling of systemic risk and endogenous contagion channels (e.g. the ECB’s BEAST model, the system-wide amplification work of Sydow et al., and the joint solvency–liquidity stress testing of Cont, Kotlicki and Valderrama to name but a few) two issues with their own implications remain:

First, “not everything that can be counted counts, and not everything that counts can be counted”. Some of the most important feedback mechanisms run through trust, reputation and market sentiment and they simply cannot be modelled exactly, at least to date. However, given their importance, it is critical to still include them, at least approximately. Agent-based models, often frowned upon by the economics discipline, remain a much underappreciated tool for doing so in my humble opinion.

Second, as a corollary of the first point, capital requirements should not be derived from exercises that feature important system-wide endogenous effects. I am partly reversing my own earlier view here: in work with Rama Cont on fire sales and indirect contagion, I argued for embedding price-mediated contagion into stress tests. I stand by that for monitoring and vulnerability identification but do no longer believe such models should set capital. The reason is twofold. First, doing so presupposes a precision that contagion models cannot, as of today, deliver. Second – and the contrast with the above-proposed Cover-1 charge is precisely the point – the objection is that it would be difficult for a regulator to implement and for a bank to accept that Bank A’s capital requirement would depend on what Bank B is currently doing or holding in its portfolio. 

Reverse stress testing, by contrast, offers a very appealing avenue to explore blind spots and Knightian uncertainty systematically. With former colleagues at a large Swiss bank, we developed a reverse stress test for IRRBB capable of running several thousand scenarios on the full group balance sheet within a few hours ; work that would be recognized by the “Swiss Risk Award” in 2024 (Schaanning et al. 2023). We showed that at the micro-scale this exercise is useful: it can differentiate between two banks that on the surface look similar, with both passing the supervisory outlier test for economic value of equity (SOT EVE), while one is running a larger hidden risk. At the macro scale, the same machinery can help answer a different question: a systematic reverse stress test across the largest institutions would reveal tothe macroprudential regulator whether “the worst” scenario is the same for most banks or diverse across them. From a system’s perspective, it is preferable that not everybody gets into trouble at the same time.

Deriving microprudential capital requirements from supervisory stress tests

All happy banks are the same, each unhappy bank is unhappy in its own way. 

While it is somewhat obvious that most, if not all, banks are vulnerable to GDP declines, falling house prices and rising unemployment, it is much less clear whether banks are vulnerable to increases or to decreases in other financial risk factors. ECB analysis from 2023 shows that they are vulnerable to both (Schaanning (2024)). From a net interest income perspective, 88 out of 102 banks were exposed to falling rates and 14 to rising rates; from an economic-value perspective, 26 out of 102 were exposed to falling rates and 76 to rising rates. The same argument applies to credit spreads, depending on whether and how banks hedge their bond portfolios and whether they are net sellers or buyers of protection. A single scenario can, by definition, only shock each risk factor in a single direction and will therefore inevitably shock many banks’ risk factors in the beneficial direction. I therefore strongly support Til Schuermann’s proposal of using at least two to three scenarios and averaging the worst outcomes in order to arrive at a fair picture of the risk landscape.

The second point concerns model calibration. Some economies, Scandinavia being a clear example, have not experienced a severe banking or macroeconomic crisis for a long time. During the Nordic banking crises in the late 80s and early 90s, annual loan losses peaked at roughly 6–7.5% of total lending (Englund (1999) and Moe, Solheim and Vale (2004)), whereas over the last decade most banks have remained at low double-digit basis points for loan losses. Models calibrated to the last one or two decades of dependence between GDP, unemployment and loan losses are therefore prone to (severely) underestimating credit losses whenever no credit crisis features in the calibration sample. This may be a third explanation for David Aikman’s question as to why ever increasingly severe stress tests keep returning comforting results of resilience.

And what about AI?

At the risk of writing a paragraph that will be outdated within a few weeks: generative AI, including generative adversarial networks (GANs), has proven remarkably effective at generating scenarios that replicate stylised market facts or that target a given portfolio’s vulnerabilities (Cont et al. 2025). Impressive indeed, but with an important caveat. These approaches still fit probability distributions to historical data. If X and Y have always been negatively correlated in the past, but for good geopolitical reasons, say, may become positively correlated in the future, these models will assign vanishingly small probabilities to positively correlated scenarios between X and Y. Overcoming this conceptual hurdle would probably require architectures capable of structural, causal reasoning about regime change – “world models”, in the current vernacular. Until then, I postulate, it will take a human risk manager with a critical brain to generate the «right bad» scenario.

Conclusion

I always pass on good advice. It is the only thing to do with it. It is never of any use to oneself. – Oscar Wilde 

Stress testing is as much art as science: it requires genuine intellectual curiosity to push boundaries, question the status quo and improve on what is missing. The stress tests conducted during and after the global financial crisis were immensely valuable in fostering financial stability. To stay relevant, however, they must evolve – in true systems-thinking fashion – along with the system itself. And they should develop, as Pedro Duarte Neves argues, into a range of tools rather than a single instrument. The range I propose runs along one dividing line :« price what is yours, explore what is everyone’s » and consists of three elements: (i) a new Cover-1 charge for the largest contingent exposure, net of the margin held, to internalise first-party endogenous risk; (ii) a “roughly right” approach to system-wide endogenous models, embraced for the exploration of vulnerabilities while letting go of deriving capital requirements from them; and (iii) where supervisory stress tests continue to set capital, multiple scenarios and a critical review of banks’ internal models through a stress-testing lens.

Endnotes

1. Reflecting the current global regulatory- and financial climate with its limited appetite for strengthening capital requirements, a first step could be to require that the counterparties creating the largest contingent liquidation risks be identified, the exposures be quantified, and the risk monitored.
2. Basel does extend the MPOR in rough ways – for example, doubling it for netting sets exceeding 5,000 trades or containing illiquid collateral or hard-to-replace trades, and extending it following margin disputes. These are binary triggers, however, not charges calibrated to the size of a position relative to the underlying market.
3. « Not even wrong » in Pauli’s words.

References

Basel Committee on Banking Supervision (2024), Guidelines for counterparty credit risk management, Bank for International Settlements.

Budnik, K., et al. (2023), “BEAST: A model for the assessment of system-wide risks and macroprudential policies”, ECB Working Paper.

Cont, R. and Kotlicki, A. and Valderrama, L. (2020), “Liquidity at risk: Joint stress testing of solvency and liquidity”, Journal of Banking & Finance, 118.

Cont R., Cucuringu M., Xu R., and Zhang, C. (2025) Tail-GAN: Learning to Simulate Tail Risk Scenarios. Management Science 72(4):2917-2936.

Cont, R. and Schaanning, E. (2017), “Fire sales, indirect contagion and systemic stress testing”, Norges Bank Working Paper 2/2017.

Cont, R. and Wagalath, L. (2016), “Risk management for whales”, Risk.net

Danielsson, J. and Shin, H.S. (2003), “Endogenous risk”, in Modern Risk Management: A History, Risk Books.

Englund, P. (1999), “The Swedish banking crisis: Roots and consequences”, Oxford Review of Economic Policy, 15(3).

Haldane, A.G. and May, R.M. (2011), “Systemic risk in banking ecosystems”, Nature, 469, 351–355.

Moe, T.G., Solheim, J.A. and Vale, B. (eds.) (2004), The Norwegian Banking Crisis, Norges Bank Occasional Paper No 33.

Schaanning, E., Hardy, C., Nunez, F., Stepanyan, A. (2023), Finding the Blind Spots Before It's Too Late: A (Reverse) Stress Testing Approach for Asset Liability Management, SSRN working paper. 

Schaanning, E.(2024),  Asset Liability Management and Interest Rate Risk in the Banking Book: an overview, SSRN working paper.

Sydow, M., et al. (2021), “Shock amplification in an interconnected financial system of banks and investment funds”, ECB Working Paper No 2581.

Disclaimer 

The views expressed are my own and should not be reported as representing the views of Leonteq or any regulatory or commercial institutionsthat I have previously been affiliated with. All errors are my own.

To comment, please go to the discussion forum