Central counterparties (CCPs) sit at the heart of modern markets, with margin representing the key mechanism for ensuring trade continuity even under stress market conditions. The quality of those margin calls depends, ultimately, on the quality of the risk model behind them. So it is worth asking, every so often, whether the models CCPs rely on are evolving at the same pace as the portfolios they are used at. The answer, we think, is: not quite as one might have expected.
The portfolios have changed. The models, less so.
A decade or two ago, cleared portfolios were typically segmented into economic groups - equity index futures here, energy futures there - and each bucket was margined more or less in isolation. With mergers between exchanges, interoperability between CCPs, the broadening of clearing mandates and the rise of cross-asset trading strategies, that picture has changed. Today’s cleared portfolios are routinely multi-asset and multi-product. They contain spread positions, basis trades, options structures and combinations of long and short exposures whose risk does not reduce neatly to the sum of its parts.
The models have evolved too, but at a different pace. According to the European Central Bank’s 2023 stocktake by Boudiaf and co-authors, of 49 CCPs and clearing services in Europe, 23 were still using SPAN as their main risk model and 20 some form of Value-at-Risk (VaR) or Conditional VaR. These are good models - transparent, computationally tractable and well-understood. But they all share, in different ways, the same simplification: to make the maths work for large portfolios, they reduce the multivariate world of risk factors to a single dimension, usually the portfolio’s profit-and-loss distribution, and read the risk metric off that.
That simplification is normally harmless. For a portfolio that is long everything, the loss distribution captures the joint behaviour of the underlyings perfectly well. But it can become a problem the moment a portfolio stops being directional.
Consider two assets whose returns are jointly normally distributed and positively correlated. The probability that both fall by some amount is not the same as the probability that the difference between the two asset returns falls by the same amount. For a meaningful subset of the statistical domain, the joint probability of the two returns is higher than the probability of an equivalent move in the spread, i.e. the sum of the returns. A traditional VaR computed on the PnL of the spread will therefore quietly understate the risk of the position. That is a feature of the geometry, not of any particular model.
This understatement is small for benign distributions and well-behaved correlations. It is not small when correlations are unstable, when tails are fat, or when the position is concentrated in spreads on the most volatile assets - exactly when the risk manager most wants the model to be working.
A different question, asked at a different layer
The starting point for what we call the expanded Value-at-Risk (eVaR) is to ask the risk question one layer further up. Rather than projecting the portfolio onto a single dimension and reading risk off that projection, eVaR works directly in the risk-factor space. For a given confidence level, it asks: across all combinations of risk-factor moves consistent with that confidence level, which one produces the largest loss for this portfolio?
Phrased that way, the problem is an optimisation. The objective is the portfolio’s loss. The constraint is that the joint scenario sits within the chosen confidence region of the risk-factor distribution. Solve it and the answer is, by construction, the worst loss the portfolio can suffer at that confidence level under that distribution.
Two consequences follow. First, the metric is portfolio-specific. Most risk frameworks generate a single set of stress scenarios, or a single Monte Carlo cloud, and apply it to every portfolio. eVaR inverts that logic: each portfolio defines its own worst case, because each portfolio cares about a different region of the joint distribution. A spread position cares about correlated moves; an option straddle cares about the centre; a directional book cares about the tail. The optimiser searches the right region in each case.
Second, the metric is computationally lighter than it sounds. The naive way to capture the same effect with a Monte Carlo VaR is to keep simulating until enough scenarios populate the relevant slice of the joint distribution. For non-directional portfolios in higher dimensions, that can be a lot of scenarios. eVaR substitutes targeted optimisation for brute-force sampling.
To test the idea, we calibrated eVaR (under a deliberately simple multivariate Gaussian assumption for risk-factor returns) on five exchange-traded futures contracts: Brent oil, MSCI World IT, MSCI Brazil, FTSE 250 and FTSE 100. The window runs from January 2020 to April 2025 and includes the COVID-19 shock, the start of the Russia-Ukraine conflict, and the April 2025 “liberation day” episode. We compared eVaR against four standard implementations on six different long-short portfolios: parametric Gaussian VaR, Monte Carlo VaR, Historical Simulation VaR and Filtered Historical Simulation VaR. All metrics were computed at 99 percent confidence over a one-day horizon, with a 250-day look-back.
The pattern is consistent. On portfolios containing spread positions on the most volatile assets, eVaR is the most conservative metric the great majority of the time, and it passes back-testing under the Basel Traffic Light System where the other Gaussian-based methods do not. On the 5-year window, those portfolios produce eVaR exceedance counts of 8 and 7 — well inside the Basel threshold of 27.
The signal is clear: when the portfolio’s worst case lives somewhere other than the simple tail of the PnL distribution, defining the risk metric in the risk-factor space pays off. And it does so even when the underlying distributional assumption is intentionally crude.
Margin is a contract between the clearing house and its members, but in aggregate it is also a piece of financial-stability infrastructure. Under-margining concentrated, non-directional positions on volatile assets is one of the recognisable patterns behind clearing-member-driven losses. Any tool that helps a CCP identify those positions earlier, and price them more honestly, is worth taking seriously.
eVaR is not a wholesale replacement for the existing toolkit. It can be deployed alongside parametric VaR, Monte Carlo or historical methods, sharpening the estimate where the existing approach is most likely to under-read the risk. Used that way, it gives CCP risk managers a practical, computationally tractable way to look harder at the parts of their books that have grown the fastest - the spreads, the basis trades, the cross-asset combinations - without having to rebuild the rest of the model stack to do it.
The portfolios CCPs clear today are not the portfolios CCPs cleared a decade ago. The risk metrics deserve to keep up.