The contracts will pay. That is the problem.

Elliot Hurdiss, CTO, FINBOURNE

The FT reported on 20 October, citing Morgan Stanley, that weekly trading in credit default swaps on Oracle, SpaceX, Alphabet, Amazon, Meta, Broadcom and Nvidia averaged over $9bn in September. In July it was under $2bn. Oracle five-year protection was 322bp on Monday against 144bp at the start of the year (a year in which Oracle has borrowed like it means it). CoreWeave is around 1,090bp, which for anyone wanting to insure $10m of its bonds means $1.82m up front and then $125,000 a quarter (at which point it is less a hedge than a forecast).

So. Let’s say the bottom does fall out of the AI trade. The obvious question is who has written all of this protection, and whether they can pay. It was the right question about AIG in 2008. It is less obviously the right one now.

What actually sank AIG

AIG had sold several hundred billion dollars of protection under credit support annexes (CSAs) that, because AIG was so highly rated, asked it for very little collateral relative to the exposure. The marks went against it from the middle of 2007 and the calls built up over a year (it had posted around $16.5bn by mid-2008, much of it disputed along the way). What finished it was the downgrade in September 2008. The thresholds in those CSAs were tied to its rating, so when the rating went the thresholds went with it, and $14.5bn more was demanded almost overnight. AIG never defaulted on a CDS. It ran out of collateral.

What changed after 2008

Most of what the market built afterwards was aimed at avoiding exactly that. Coupons were standardised so contracts could be netted and compressed. Index CDS went into mandatory clearing. The uncleared margin rules covered the rest. It worked. A protection seller today pays variation margin (VM) every day as spreads move and, if it is large enough to fall under the initial margin (IM) rules, posts IM from the day after the trade.

The sellers are different too. The dealer is the name on the other side of the ticket, but dealers keep single-name books close to flat because the capital charges for warehousing credit are punitive, so the risk is passed on to hedge funds running relative-value and index arbitrage, structured credit vehicles that have sold tranche protection, and real-money accounts selling protection for the carry. The risk is collateralised and spread out, but a good part of it is held by levered accounts.

The margining did not get rid of the credit risk. It just changed when and how the seller pays for it. A seller still loses money when spreads widen, but loses it in daily cash instalments, and each instalment must be funded that day, in cash or eligible securities, by selling or pledging something.

When spreads drift wider over a year that is fine, and it is what the system was built for. When several names gap in the same week, it is less fine. Initial margin models are calibrated on history, and there is no AI credit event in the history, so IM will probably turn out to have been set too low. Variation margin calls land on every seller on the same day. And what the sellers sell to meet them is what everyone else is selling.

The March 2020 precedent

The nearest precedent is March 2020. COVID and an oil price war pushed volatility up, and the clearing houses responded the way their models told them to. CME raised margins on Treasury futures several times during the month. Across all clearing houses, initial margin rose by something like $300bn in March, according to the regulators’ review afterwards, and daily variation margin calls went from an average of about $25bn in February to $140bn on 9 March.

Hedge funds in the Treasury basis trade got those calls, found the repo that funded the trade had dried up, and sold the Treasuries. So did a lot of other people who needed dollars. Somewhat remarkably, Treasuries fell during a flight to safety, which is not supposed to happen, and the Fed ended up buying around $75bn of them a day for a period, on top of a new dealer facility and much larger repo operations. Same story in 2022 with our now infamous “mini-budget”, with the Bank of England in the Fed’s role.

So the worry is not a seller refusing to pay. It is all of them paying, every day, into a falling market, when some of what they are selling is also in your portfolio.

How it reaches your fund

That is one way it reaches you. The other is your own margin book, and for that you do not need to own any of these names. When AI credit gaps, volatility goes up everywhere and every margin model reprices with it. Initial margin rises on positions that have nothing to do with AI, clearing houses and prime brokers tighten what they will accept, and the CDX hedge put on against exactly this starts generating calls of its own. The cash is due that day, at the same moment everyone else’s is.

How not to be a forced seller

The call is not what hurts. Being a forced seller to meet it is. The fund that avoids the hit is the one that already knows what it will owe and what it can deliver without selling anything. That takes four things.

Calculate your own margin. Every agreement, every evening, VM and IM, from positions reconciled against the counterparty that day, against the terms actually in the document. If the first time the number appears is on the counterparty’s demand, the fund is a day behind before it starts, and it is arguing about someone else’s number.

Know the inventory. What is held, in which account, whether it is eligible under each schedule and whether it is already pledged somewhere else. That has to come out of the holdings, not a balance someone typed in. And know which agreements give the other side discretion over haircuts, because that is the part of the book the cash buffer needs to be sized for.

Run the stress through to the calls. Not a portfolio VaR. Tomorrow’s margin, by agreement, over the next ten days, with spreads gapping, equities falling and a counterparty downgraded, set against what is actually free to post. The gap between stressed calls and free inventory is the number that matters, and most funds have never seen it.

Decide in advance what gets delivered. On a same-day call cash at T+0 beats a bond settling T+1 whatever it yields. Rank what is eligible by what it costs to give up, have the substitutions ready, and do not make the choice under pressure, because selling the wrong asset to raise cash is how a margin call becomes a realised loss.

Why most operating models fall short

Most operating models cannot do this on the night. Positions arrive on the custodian statement overnight, the demand arrives in the morning as a PDF with the counterparty’s number on it, and reconciliation happens after the call has been agreed. The stress-to-calls piece mostly does not happen at all, because it needs valuation, risk and collateral working off the same data at the same time, and in most shops they are three systems and a spreadsheet.

That is the gap FINBOURNE can close. LUSID puts risk scenarios, collateral agreements, the margin calculation and a liquidity forecast on one reconciled dataset, so a stress run in the risk engine becomes a stressed margin forecast for each agreement, set against what is actually free to post.

The funds that come through a margin spiral in reasonable shape will be the ones that knew on Tuesday night what Wednesday’s calls were going to be, what they would meet them with, and had already moved the cash. Calling the AI bust correctly will have had little to do with it.

To see how FINBOURNE turns a stress scenario into a margin forecast for each agreement, ask us for a demonstration.