When markets move into territory a risk model says is impossible, the model usually fails in one of two ways. It stops running, or it keeps running and quietly produces nonsense. The first failure is the better one, because at least you know. In my view, the real lesson for financial institutions is that "impossible" is a design choice. It is a constraint someone wrote into the model at some point, often for good reasons. Every one of those constraints should be found, documented and tested before the market tests it for you. Then they need a deliberate decision about which extremes to allow and which inputs to reject.
Assumption → Breach → Fix → Validate
Most firms meet these four stages in the wrong order. They discover the assumption only when it is breached, fix it under pressure overnight, and validate the fix afterwards, if at all. I would reverse that. Name the assumptions while things are calm, rehearse the breach, decide in advance what the fix looks like, and validate both the extreme behaviour and the guardrails against bad data.
Assumption: every model has a floor and a ceiling somewhere
Pricing and risk models carry assumptions about the world that rarely appear in a governance document. Some are mathematical. A lognormal rate model cannot represent a negative rate, because the logarithm of a negative number does not exist. Some are coded as sanity checks: reject a price below zero, cap a volatility, refuse a correlation outside a range. Some sit in the scenario engine, where a fixed bump is applied to every curve regardless of where the curve starts.
Each of these is reasonable when it is written. Together they form an invisible map of what the institution believes cannot happen. The trouble is that nobody owns the map as a whole.
The underlying distributions carry assumptions too. In my conversation with Dominic O'Kane, who built risk and pricing systems at Salomon Brothers and Lehman before turning his teaching code into the open-source FinancePy library, he described the standard Black-Scholes world of geometric Brownian motion simply. Daily returns come from a bell curve, and the process has no big jumps. His test for telling the model from the market was blunt: "If you find a big jump, then that's the real world." That is why large banks build a whole zoo of jump models on top of the standard one. The base assumption is known to be wrong in exactly the moments that matter most.
The same gap can appear in a carefully built library. When I raised jump-to-default on a credit default swap, the risk that a credit is fine one day and in default the next, Dominic acknowledged that FinancePy's CDS covers interest rate, spread and recovery risk, but not loss on default. He noted it is a simple calculation to add. I do not raise this as a criticism. I raise it because it is typical. A model can be rigorous about the sensitivities it computes and still be silent about the scenario a risk committee most wants to see. Silence is an assumption too.
Breach: the night the impossible arrived
The clearest account of a breach I have heard comes from Dominic's early career. In 1994, Salomon's risk system ran perturbation scenarios overnight, bumping rates down by 25 basis points. Japanese rates were sitting at around 20 basis points. Apply the bump and the scenario rate goes negative. The models had no room for that, so the run simply stopped. In his words, they had to do "an emergency fix to allow negative interest rates." At the time, nobody expected to see the problem again.
A few points in that story matter to any institution today.
First, the market did not breach the assumption. The scenario did. Actual Japanese rates were still positive. The firm's own stress process produced a state its models could not handle. Many breaches start this way. A fixed-size shock, applied mechanically, drives a near-zero quantity through zero. If your stress framework applies the same absolute bump to every curve, check what it does to curves that are already close to a boundary.
Second, the failure was loud. The system crashed and someone had to fix it before traders arrived. That is uncomfortable, but it is honest. The worse version is a model that floors the rate at zero without telling anyone and reports a risk number that looks normal.
Third, "we will never see this again" is not a control. About a decade ago, Swiss rates and others went negative. Interest rate option models, including caps, floors and swaptions, had to be rebuilt or reparameterised to cope. Rates that had been an anomaly in one overnight run became a sustained regime across several markets.
The breach can come from either direction. Oil futures went negative around COVID. LME nickel went up by a factor of four in three or four trading days. One market went radically below what models assumed and the other went radically above it. A model that tolerates one of those cases may still break on the other. Upper bounds, volatility caps and margin assumptions deserve the same scrutiny as zero floors.
Fix: allow the extreme without accepting the error
The obvious fix is to remove the constraints. Let rates go negative, let prices go wherever they go, let the model run. That solves one problem and creates another. Dominic called it a double-edged sword: if someone enters "a negative price for a futures contract, you kind of want to think that that's a mistake." Most extreme inputs in practice are not market events. They are typos. His example was a 5% coupon entered as 5 instead of 0.05. That is an error of two orders of magnitude, and it will flow through every downstream number if nothing stops it.
So the fix is not to drop the guardrails. It is to separate two kinds of check that most systems blur together:
- Domain constraints are what the mathematics can handle. These should be as wide as the real world can plausibly go. If a rate can be negative in any market you trade, the model should handle negative rates. That might mean a shifted or normal model, or simply removing a hard floor.
- Plausibility checks are what the data probably should look like. These should be narrower and soft. Outside the expected range, the system should flag the value, ask "are you sure," and record who confirmed it. It should not refuse to run.
Hard stops belong where the input is incoherent. Soft warnings belong where the input is unusual. A negative futures price in April 2020 was unusual but real. A coupon of 500% on a vanilla bond is almost certainly a keying error. Your systems should be able to tell those two cases apart, and the people overriding a warning should leave a trail.
Validate: test the edges, not just the middle
Model validation tends to concentrate on the region where the book actually sits: normal rates, normal volatilities, normal correlations. That is where most of the money is today. It is not where the surprises come from. For the unhedgeable inputs, I have written separately about how validation stresses parameters and holds reserves. Boundary validation is a different discipline. It asks whether the model behaves sensibly when an input crosses a line the developers assumed nobody would cross.
I see a newer reason this matters. AI coding tools now let analysts and risk managers build models and interfaces on top of libraries they did not write. That is a real gain in who can build financial software. But generated code inherits whatever assumptions the prompt never mentioned. A tool asked to "price an interest rate cap" may well reach for a lognormal model and a zero floor, because that is what most textbook examples do. The person directing the tool has to know enough about the market to say: rates can be negative here, prices can jump, defaults happen overnight. Domain judgement is becoming the main control, because the code itself is cheap.
A checklist for risk and model governance leaders
- Inventory the impossibles. For each production model, list every floor, cap, log transform, positivity check and range restriction, whether in the maths, the code or the input layer.
- Run boundary scenarios on purpose. Push near-zero quantities through zero and push prices to several multiples of today's level. Confirm what happens in each case: the model runs, fails loudly or fails silently.
- Audit your shock design. Replace or supplement fixed absolute bumps with shocks that respect where each curve starts, and confirm the engine can represent the resulting states.
- Split hard stops from soft warnings. Document which input checks are domain limits and which are plausibility flags. Make overrides logged and attributable.
- Ask what the model is silent about. Jump-to-default, gap risk and liquidity cliffs often sit outside the reported sensitivities. Make the omissions explicit to the people reading the risk report.
- Validate AI-assisted builds for embedded assumptions. Review generated pricing code specifically for textbook defaults that do not hold in your markets.
- Retire the phrase "that can't happen." Replace it with "that is outside our current range, and here is what we do when it isn't."
Markets have already shown they will produce negative rates, negative prices and fourfold moves in a week. The institutions that handle the next one calmly will be those that decided in advance how their models should respond.
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.