Banks hold back trading profits on an unhedgeable model risk by booking a reserve. Model validation estimates the uncertain input, stresses it across a plausible range, and withholds the gap between the trader's headline profit and the worst-case value. That reserve stays off the trader's P&L until the risk is hedged, offset or resolved by the market, and it is released only as the sensitivity actually falls. The profit is real on paper. The bank simply declines to count it until it has reason to believe it.
Estimate → Stress → Reserve → Release
Every unhedgeable position should move through those four stages. Each one is a control point, and each is a place where the institution, the trader and the validation team can disagree. When I look at how a firm handles model uncertainty, I want to know who owns each stage and what evidence moves a dollar from one stage to the next.
Why some risks fall outside the hedge
To see why reserves exist, start with the risks that do not need them.
Most derivatives pricing rests on the logic Black, Scholes and Merton established in the early 1970s. A bank sells a client a call option on, say, IBM. The payoff is nonlinear: the client exercises above the strike and walks away below it. The bank does not need a view on where IBM is going. It buys or sells enough of the underlying stock each day that any move in the share price is offset by an equal and opposite move in the option's value. Done frequently enough, this delta hedging leaves the book almost indifferent to direction. That is why two traders with opposite views on the stock can still agree on what the option is worth. The price comes from the cost of hedging, not from a forecast.
The whole arrangement depends on one assumption. In our conversation, Dominic O'Kane stated it plainly: the Black-Scholes world "assumes that for every risk, you can find an instrument that you can use to hedge it." Dominic spent years building credit derivative models at Lehman Brothers and now maintains FinancePy, an open-source pricing library. He has lived on both sides of that assumption.
Credit markets in 2002 and 2003 pushed hard against it. Spreads were tight, and investors wanted more yield. One answer was the first-to-default basket. Instead of taking the default risk of one name, the investor took exposure to five names and lost on the first one to default. The price of that trade depends heavily on default correlation, meaning how likely the five credits are to fail together rather than separately. No liquid instrument let a dealer hedge that correlation directly. The Gaussian copula became the standard tool for pricing it. I cover why that particular model won out in a separate piece on speed versus sophistication. Whatever model the desk used, the correlation input remained an estimate nobody could lock in.
This is where reserves come in.
Walking through the four stages
Estimate
The desk prices the trade using a correlation assumption. In the example Dominic gave, the trade is marked at 25% correlation, while validation's own central view sits nearer 20%. Already the estimate is a negotiated number. The desk has every incentive to choose the input that maximises day-one profit. Validation's job is to say how wrong that input could be.
Stress
Validation sets a plausible range, in this case roughly 15% to 35%, and reprices the trade at both ends. The trader believes the deal has made $500,000. At the unfavourable extreme, it is worth only about $250,000.
The method itself is simple. What matters is that the range is defensible, documented and set by someone other than the person who gets paid on the outcome.
Reserve
The difference becomes a reserve. Dominic described the conversation with the trader bluntly: you cannot take that $500,000 into your P&L "and use it to get a bigger bonus this year." Half the reported profit is held back until the market gives better information about where correlation really sits.
I think executives underestimate how much institutional judgement sits inside that one line. A reserve is the bank stating that the model's output is not the same thing as earned money. It is a deliberate gap between accounting value and economic confidence.
Release
This is the stage I find most revealing. The trader does not simply wait. He goes back to clients and sells second-, third-, fourth- and fifth-to-default baskets on the same names. Each of those trades carries correlation sensitivity in the opposite direction to the first-to-default. Sell the full ladder and the correlation exposure largely cancels. With each new trade, the trader returns to validation, shows that sensitivity has fallen, and draws reserve back into P&L, perhaps $50,000 at a time.
The mechanism is sound. A reserve that never releases would punish genuine risk reduction. But release also connects directly to compensation, and that is where governance has to be strongest.
The part the textbooks leave out
When Dominic finished the example, I pointed out that nobody talked about bonus pools rising as correlation reserves were released. He agreed, and added that you never read about this in John Hull's book.
He is right, and the omission matters. Reserve release is not just a risk mechanism. It shapes behaviour. A trader with $250,000 locked up has a strong reason to originate the offsetting trades, and that can be healthy. It can also push a desk to sell structures to clients mainly because they unlock reserve, not because they suit the client. It can also lead a desk to argue that a hedge is cleaner than it really is. When a release is approved, the validation team is effectively approving part of someone's bonus.
For a bank, then, the question is not whether reserves exist. It is whether the release criteria are as rigorous as the original stress. I have seen firms put real effort into setting reserves and much less into checking whether the offsetting positions genuinely neutralise the risk under the same range of assumptions.
How the reserve function changed after 2008
Before the financial crisis, this kind of reserving relied heavily on desk-level judgement and negotiation. Since then, product innovation has slowed, proprietary trading desks have largely disappeared, and the regulatory framework around trading books has become far more prescriptive. Dominic observed that the role of the quant now includes a regulatory function, because how reserves are calculated has become much more quantitative than it was.
I see that as progress, with one caveat. A more formal reserve methodology is easier to audit and harder to game. But formalisation can also create false comfort. A reserve calculated to three decimal places is still built on a range somebody chose. The discipline that matters most is still the honest question at the stress stage: how wrong could this input be, and who decided?
What this means beyond the sell side
Asset managers and other buy-side institutions should care about this even if they never book a reserve themselves. When a dealer prices a correlation-sensitive or otherwise hard-to-hedge product, the cost of the reserve and the uncertainty behind it are built into the price the client pays. A buyer who understands which inputs are unhedgeable is in a better position to challenge that price, ask for sensitivities, and judge whether an exotic structure is worth the premium.
The same logic applies internally. Any firm that marks positions with model inputs it cannot observe or hedge, such as private credit, structured products or illiquid volatility, faces the same choice. It can recognise the model's number as profit, or it can hold some of it back until the uncertainty narrows.
A checklist for executives
- Inventory the unhedgeable inputs. For each material model, list the parameters that cannot be observed or hedged in the market.
- Separate estimate from stress. The person proposing the mark should not set the plausible range.
- Document the range and its basis. Record why the bounds sit where they do, and revisit them when markets shift.
- Hold release to the same standard as reserve. Require evidence that offsetting trades reduce sensitivity across the full stressed range, not just at the central estimate.
- Make the compensation link visible. Report reserve releases alongside the P&L and bonus accruals they feed, so senior management sees the connection.
- Ask counterparties the same questions. On the buy side, request sensitivities to unhedgeable inputs before accepting a dealer's price.
The model will always produce a number. A sound institution needs a clear, defensible process for deciding how much of that number it believes, and for showing the evidence each time it decides to believe more.
About the author
Michael McLaughlin has more than 30 years in structured finance and private credit. At Macquarie Group from 2008 to 2023, he built the credit markets business from inception into a portfolio financing platform that averaged a $5 billion portfolio and generated more than $1 billion in cumulative net revenue. He has advised Hebbia on AI research software for financial services, and builds his own finance applications with AI coding tools and open-source software.
Through Path66, Michael works with financial institutions and investment firms on asset finance, liquidity, capital raising, data-driven due diligence and the practical application of emerging technology in finance. He hosts Stack to Street, where he interviews the developers building the software that will change financial institutions. Full bio.
Source
From Physics to Wall Street, Stack to Street with Dominic O'Kane, Creator of FinancePy. Full transcript on the episode page.