Risk Calculation in an AI World, Part 1: How We Calculate Risk Today
The part of financial markets that I have obsessed over for the last few years has been clearing.
For the first decade of my career I used to focus on trading algorithms, but that has long since become a game of co-location and processing power. It also overlooks the single biggest cost for trading business and the market in general, inefficiencies in clearing and over margin.
This first in a short series on the impact of AI on this space is more of a literature review than a position piece. The first thing that looking at this forces on you is to inventory how the industry calculates risk now. Not the textbook version. The production version: the models that actually set margin at CME, Eurex and ICE, the ones that decide how many billions of collateral move between firms every day.
This one lays out the current machinery in its own frame, mostly because you cannot see what is changing until you can see what is there. Part two describes the engine we have been building, where the mathematics differs, and what AI engineering techniques make possible that was not practical before. Parts three and four are the consequences: first for liquidity in a stress, then for the people who do this work.
This is important for anyone that has to understand risk and cost in trading. A margin model is not a reporting tool. It is the price of holding risk. It decides how much capital a position consumes, which trades are worth doing, and how much liquidity gets pulled out of the market in a stress. The methodology choices inside these models are market structure, whether or not anyone outside the model validation team ever reads them.
The VaR family
Almost everything in production descends from one idea: take the current portfolio, subject it to a set of historical market moves, and read a loss number off the resulting distribution at some confidence level. Value at Risk.
The first generation was parametric. J.P. Morgan’s RiskMetrics, published openly in 1994, assumed returns follow a normal distribution, estimated volatilities and correlations from recent history, and computed the loss quantile from a formula. Two years later the Basel Committee’s Market Risk Amendment let banks compute regulatory capital from their own internal VaR models (Basel Committee on Banking Supervision, “Amendment to the Capital Accord to Incorporate Market Risks,” January 1996, https://www.bis.org/publ/bcbs24.htm), and VaR became the shared language of market risk. The normality assumption was always the known weakness. Market returns have fat tails, and a model calibrated to calm periods understates exactly the moves that matter.
The second generation dropped the formula and replayed history directly. Historical simulation reprices today’s portfolio under each of the last several hundred days of actual market moves and reads the loss distribution off the results. No distributional assumption, but a new problem in its place: the model can only know about moves that happened inside its lookback window. A two-year window in 2007 contained no crisis.
The third generation, filtered historical simulation, is where the industry mostly lives today. FHS keeps the replay of history but rescales it. Each historical return is divided by the volatility that prevailed at the time (estimated with a GARCH-type filter) and multiplied by current volatility. A move from a calm month in 2019 gets amplified if markets are stressed today; a panic day from 2008 gets damped if today is quiet. The effect is that thirty years of history become usable at today’s volatility level. It is a genuinely useful fix, and it is the common core of the three big production models. It is an answer to computational constraints.
What actually runs at the clearing houses
The production implementations are worth seeing in detail.
CME’s SPAN 2 runs filtered historical VaR over a lookback window of at least ten years, with observations weighted across time buckets and the March 2020 COVID moves deliberately retained even as they age out of shorter windows. Margin is taken at 99% confidence. Anti-procyclicality is bolted on by rule: margin may not rise more than 35% inside 30 days, and a floor is set at the 25th percentile of the trailing ten years of margin levels. Its predecessor, the original SPAN of 1988, scanned a portfolio against sixteen prescribed scenarios and took the worst; the replacement of sixteen scenarios by a decade of filtered history is itself a compact history of what cheaper compute did to risk modelling.
Eurex’s PRISMA computes market-risk margin as the worst of three numbers: filtered historical VaR, a VaR computed over designated stress periods, and an Event Risk charge built from hypothetical jump scenarios for risks with no useful history, such as a sudden dividend cut or a crypto de-peg. (Eurex publishes its methodology in brochures and circulars; the description here is reconstructed from those public materials.) PRISMA’s signature discipline is about netting rather than the loss model: offsets between products are granted only within Liquidation Groups, sets of products that could genuinely be auctioned together if a member defaulted. Sensitivities cancelling on paper is not enough; the default manager has to be able to liquidate the hedge as one package. Different Liquidation Group Splits carry different margin periods of risk, two days for listed fixed income against five days for OTC interest rate derivatives, reflecting how long a default auction would actually take. And Eurex publishes Transparency Enabler files that let members replicate their own margin calculation offsite, which tells you what the industry considers the gold standard of model accountability: the customer can reproduce the number.
ICE’s risk model is the most instructive of the three, because ICE wrote its reasoning down in a patent (US Patent 10,922,755 B2, published as US2020/0202443A1, https://patents.google.com/patent/US10922755B2). The engine described there runs roughly 1,250 EWMA-filtered historical scenarios with full revaluation of every instrument under every scenario, takes the 99% quantile, and scales for the margin period of risk. For the interest-rate futures silo alone the patent discloses 476 risk factors, and the volatility filtering across those factors works out to something like three quarters of a billion EWMA computations per margin run, before any instrument is priced. On top of the core VaR sit the add-ons: a liquidity risk charge for concentrated positions, a correlation stress charge that penalises over-reliance on cross-asset diversification, and a hard cap that limits diversification benefit to a fixed fraction of gross margin
One paragraph of that patent deserves to be quoted, because it is the fork in the road this whole series is about. ICE explains that a full valuation approach was selected rather than local sensitivities because “approaches like delta-normal capture linear risk exposures only.” In other words: sensitivity-based methods cannot see the curvature in an options book, so we will reprice everything under everything, and pay the compute bill. Hold that thought for part two, because the engine we have been building is a bet that the trade-off ICE describes no longer exists.
The sensitivity route
There is a second lineage, and it matters because it owns the largest uncleared market in the world.
ISDA’s Standard Initial Margin Model, SIMM, live since 2016 for margining derivatives that do not clear, works not from historical scenarios but from risk sensitivities. Each counterparty computes its portfolio’s deltas, vegas and curvature exposures across six risk classes in a standard format (the Common Risk Interchange Format, CRIF), multiplies each by a prescribed risk weight, and aggregates with prescribed correlation matrices through a nested square-root formula. Risk weights are calibrated to a 99% confidence over a ten-day close-out period, updated once a year through an ISDA governance process, and applied identically by everyone.
SIMM’s design goal was never maximum accuracy. The netting is poor by any other clearing metric than pure bi-lateral. It was about agreement. Two counterparties who each compute the other’s margin need to reconcile to the same number, dispute line by line when they differ, and do this across thousands of bilateral relationships without a daily negotiation. A sensitivity-based formula with published weights does that in a way a full historical simulation never could, because the inputs are a few hundred standardised numbers per portfolio rather than a decade of proprietary scenario data. The trade is explicit: SIMM is deliberately conservative and deliberately coarse, and everyone accepts the overshoot in margin as the price of a number both sides can compute. The important point here is that each entity can’t be running a full historical simulation in this bi-lateral space.
Notice what the two lineages have in common despite their differences. Both fix the risk representation in advance: a scenario set on one side, a sensitivity grid on the other. Both push everything difficult into calibration, which is periodic, governed and slow. And both were shaped by a constraint that had nothing to do with mathematics: the result had to be explainable to, and checkable by, the humans on the other side of it.
The tail, and the patches
Everything above shares one deep limitation. The models learn from history, and the loss that breaks a clearing house is by construction not in the sample yet.
The mathematical response has been extreme value theory. EVT fits a Generalised Pareto Distribution to the exceedances beyond a high threshold, on the theorem-backed grounds that tails converge to that family almost regardless of the underlying distribution. It gives you a disciplined way to extrapolate beyond your data instead of pretending your worst observed day is the worst possible day. In production it appears mostly at the margins of the big models, informing stress calibrations and tail scalings rather than driving the headline number. EVT in discussion is usually a sign that the reader knows what they are talking about, in the way that portfolio gap risk doesn’t.
The regulatory response is a stack of patches, each one sensible, each one adding a layer. Stressed VaR, introduced under Basel 2.5 in 2009 after the crisis made the problem impossible to ignore, requires banks to run their VaR model a second time against a twelve-month window of significant stress and add the result to capital (BCBS 158, https://www.bis.org/publ/bcbs158.htm). The trading book rules that followed replaced VaR with expected shortfall, calibrated to the most severe stress period observable for the portfolio. Clearing houses in Europe must implement at least one of three prescribed anti-procyclicality tools under EMIR: a 25% margin buffer that can be released in stress, a 25% weighting on stressed observations in the lookback, or a floor based on ten-year volatility (EMIR RTS 153/2013, Article 28).
The anti-procyclicality tools exist because of a finding that deserves to be more widely known outside the specialist literature. Murphy, Vasios and Vause, in a Bank of England Financial Stability Paper in 2014, established that risk-sensitive margin models are procyclical by their nature: they demand the most collateral at exactly the moment liquidity is scarcest, and every available mitigation buys stability at the cost of risk sensitivity or capital efficiency. There is no parameter setting that removes the effect. You can only choose your trade-off. ESMA’s own review after the COVID margin spikes and the LME nickel suspension reached the same conclusion: none of the three prescribed tools eliminates procyclicality.
So the state of the art in the tail is: extrapolate carefully with EVT where you can, run the model again on the worst year you have seen, hold a buffer, add a floor, add a cap, and accept that the model will still tighten the screws in a crisis because that is what risk-sensitive models do.
What the whole stack is telling you
Step back from the detail and look at the shape.
Every production model runs in batch, mostly daily, because full revaluation across a thousand scenarios is expensive and because the operational rhythm of margin calls is daily.
Every model fixes its risk representation in advance and pushes adaptation into a governed calibration cycle, quarterly or annual, because each recalibration has to pass a validation committee and, for a clearing house, a regulator.
Every model handles what it cannot represent with an add-on: a liquidity charge here, a correlation stress there, a floor, a buffer, a cap.
Conservatism is distributed across a dozen patches, each individually justified, none of them priced against the others. These compound and we extract more and more capital out of the markets to “protect”, sometimes the cause of the issues we see, lack of growth, is from the protection mechanisms themselves.
And here is the detail that tells you which of those constraints was actually binding, because the industry ran the experiment.
The compute constraint fell first, and it fell nearly twenty years ago.
Adjoint algorithmic differentiation collapsed the cost of computing full-portfolio sensitivities by orders of magnitude from around 2006 (Giles and Glasserman, “Smoking Adjoints,” Risk, January 2006), and banks spent the 2010s building GPU risk farms on top of it.
If expensive computation had been what kept risk calculation in its nightly batch, the architecture should have transformed by 2015. It did not move: margin stayed daily, models stayed governed, recalibration stayed annual.
SIMM, designed in the mid-2010s when compute was already cheap, was built deliberately simple anyway, because what its designers were rationing was not computation but agreement: the ability of thousands of counterparties to reconcile a number without a phone call.
The constraint that shaped this architecture was the cost of checking, agreeing and validating rather than the cost of computing, and that is the constraint the rest of this series is about.
Every clearing house, meanwhile, runs one model per product silo, with netting granted grudgingly across the boundaries, capped at ICE, gated by auction feasibility at Eurex.
The reasons are partly good (Eurex’s liquidatability discipline is a real insight about defaults, and we will come back to it in part two) and partly institutional: every model is a separate validation burden, a separate filing, a separate thing a committee must hold in its head.
The silos this newsletter keeps returning to, the ones that force offsetting positions to be margined twice at different venues, are not a conspiracy. They are what you get when the unit of regulatory approval is the model and models are expensive to check.
None of this is a criticism of the people who built it. The stack is coherent, battle-tested, and honest about its patches in a way vendors rarely are. It cleared the COVID shock and the 2022 rates repricing without a major CCP failure. It is what good engineering looks like when history is the only available model of the future and human verification is the scarcest resource in the system, rationed by committees, calibration cycles and model-per-silo approval.
The question this series asks is what happens when that last constraint moves.
Not “will AI predict the market,” which is the wrong question and the answer is no.
Ask instead: what has the arrival of AI actually done to quantitative finance?
The series of articles everyone expects on that subject is about speed, models drafted faster, code written faster, juniors made more productive. That has some truth and it is the small answer.
The bigger issue one is that the mathematics and the engineering that produced large language models, the geometry of high-dimensional spaces, the train-once-query-cheaply architectures, the machine-enforced verification that makes untrusted generation safe to use, amount to a different way of seeing risk itself, and are where the interesting next generation of this work is.
And almost none of it is taught to, or chosen to be studied, by the people who become quants.
The rest of the series makes that case with a build log and POC rather than a manifesto.
Part two is the engine: a clearing risk engine built the way inference systems are built, where the mathematics genuinely differs from the stack described here.
Part three takes the strongest objection head-on, that faster and sharper risk measurement makes the liquidity spirals worse, and shows why the cure lives in a different layer of the model than the measurement.
Part four returns to the question above and to the people: what the new skill set actually is, why the industry is short of it, and how much retraining stands between here and there.
AI is changing the way that we think about development, not just as a tool to write code, but how we approach complex big data problems. Risk is very much in this class, solving it the way AI has been developed rather than solving it the way machine learning was operating is the obvious way forward.
We don’t want to say we are an AI business for Risk, that would be wrong, but we are build Risk that uses the techniques from AI engineering to dramatically improve the way Risk is used in clearing to improve netting.
This first appeared on LinkedIn on 7 July 2026. If you want to comment or discuss, that’s the place.