Independent deterministic validation of market risk · for the largest portfolios and trading books.
VaR, CVaR, stress and tails computed with deterministic risk computation and reproducible hash reproducible · as an independent challenger alongside Aladdin, Bloomberg, Murex, MSCI. For large asset managers, hedge funds and banks. Verification on your data, without replacing your existing stack.
The largest institutions have the systems · but no independent deterministic control, and more compute cannot be bought
Aladdin, Bloomberg, SAS are everywhere. What’s missing is an independent, reproducible check of model risk · mapped to SR 26-2, FRTB IMA and Basel governance · and the ability to scale tail computation when buying more GPUs is physically impossible.
You can no longer buy more compute speed with money: TSMC’s advanced packaging (CoWoS) is sold out at least through 2027. Building a full independent validation function internal takes years and hundreds of FTE. That leaves a third path: an external deterministic challenger that computes faster on the same hardware and produces a reproducible trail for the regulator.
Speed, reproducibility and independence · at the same time
Monte Carlo does not scale on a frozen GPU fleet: the larger the book and the scenario set, the slower and costlier the run. And the regulator wants not just a number, but a provable, reproducible conclusion from an independent party.
An internal team computes on the same engine as the front office · that is not an independent check. The Big Four give an opinion, not a reproducible computation. In both cases: not hash reproducible and without an audit of every step.
- 01 Monte Carlo does not scale · you cannot buy more GPUs
- 02 Validation is understaffed · models grow faster than the team
- 03 Internal computation is not independent of the front office
- 04 A consultant’s opinion is not a reproducible trail for the regulator
Digital asset market crash · $19 billion in 24 hours
On 10 October 2025, more than $19 billion in positions were force liquidated within hours · the largest event of its kind in market history (CoinDesk). The cause was a combination of high leverage, liquidity that vanished instantly, and blunt automated liquidation against a sharp macro trigger. It is precisely in cascades like this that standard VaR models, calibrated on «calm» markets, systematically underestimate the tail. In our engine, tail metrics (CVaR / ES, EVT) and scenario stress are the default computation, not an emergency measure.
Archegos · concentration invisible to the prime brokers
Family office Archegos accumulated enormous positions through total return swaps so quietly that even its own prime brokers couldn’t see the aggregate exposure. The result · more than $10 billion in losses at counterparty banks (Credit Suisse, Nomura, Morgan Stanley, UBS); the Credit Suisse losses contributed to its later collapse and absorption by UBS in 2023 (InvestorLawyers). There was no independent counterparty concentration check · and that became the systemic risk.
Both the tail cascade and the hidden concentration are caught by the same metrics (CVaR/ES, EVT, concentration map Top 1 / 2 / 3, HHI) · as an independent check on top of your existing stack, before a counterparty or the regulator sees it.
An independent challenger on top of your stack · on your data
You hand over positions or the book; we return an independent, reproducible computation of VaR/CVaR, stress, tails and concentration that you reconcile against Aladdin/Bloomberg and against your realized dynamics. No system replacement, no deployment.
Data handover
Positions or the book over a secure channel. NDA before any data is received.
Independent computation
VaR/CVaR, stress, tails, concentration · deterministic, hash reproducible.
Validation packet
Challenger report + backtest (Kupiec / Christoffersen / Acerbi-Székely) + audit trail.
Monitoring
Regular recomputation, signals on divergence from the front office and building risks.
ARIN22 · an internal deterministic risk computation kernel
Institutional risk · VaR, CVaR, stress, tails · is computed with a deterministic numerical kernel, not random sampling: the same input yields the same result, reproducibly, with every computation step auditable. Every run and every test is verified through hash pinned replay on CPU and GPU.
The kernel’s core capability is not just to compute a number, but to decide whether it can be used: ready / under review / data required. Class routing between deterministic computation, a Monte Carlo challenger, and a fail closed mode · with no silent degradation of the result.
It embeds alongside Aladdin, Bloomberg, Murex, MSCI · as an independent check, not a replacement. Verification runs on your data: you reconcile the result against your own realized book dynamics.
What it computes · the full spectrum
- •Market riskVaR (Historical / Parametric / EVT / Monte Carlo), CVaR/ES, risk contribution, volatility, Sharpe, MaxDD, Calmar.
- •Tail and extreme scenariosEVT/POT, FHS, hybrid tail, deterministic deep tail functional, levels 99 / 99.5 / 99.9.
- •Monte Carloportfolio MC, deterministic MC equivalent, hybrid mode, MC VaR/CVaR, stability controls.
- •Stress scenarioshorizons 30/90/180/360 days; historical (COVID, 2008, 2022, Black Monday, inflation shock); thematic; a catalogue of 28 scenarios.
- •Factors and dependenciesprincipal components and explained variance, factor exposures (26 factors), contribution/residual, correlations, covariance with shrinkage.
- •Fixed incomeDV01, duration, convexity, rate shocks (+25/+50/+100/minus 50 bp) and spread shocks (+50/+100/+250 bp).
- •ConcentrationTop 1 / 2 / 3, HHI, effective number of assets, diversification ratio, limit breach history.
- •Model validationKupiec, Christoffersen, Acerbi-Székely (ES), EVT fit check, PIT (power aware), shadow ledger of realized outcomes.
- •Technical labsH100 scale, D=200 dimension, 10M paths per case, option Greeks, implied volatility surface, FRTB curvature · specific numerical methods under NDA.
Hardware scarcity cannot be bought around · software is what’s left
Advanced chip packaging capacity (CoWoS) at TSMC is sold out at least through 2027, and leading edge 2nm nodes are contracted years ahead. TSMC’s CEO puts it plainly: CoWoS capacity is «very tight and sold out… through all of 2026.» In the coming years, the largest institutions physically cannot solve the shortage of compute speed by buying more GPUs · what’s left is software that squeezes more out of already installed hardware. ARIN22 computes CVaR in milliseconds on the same hardware where classical Monte Carlo takes orders of magnitude longer: an answer to a shortage that will only deepen, not a hypothetical advantage.
The first to hit this constraint are those with the biggest books: large asset managers, hedge funds and banks’ trading books · both buy side and sell side. That’s exactly where we aim.
Same input. Same hash. Every time.
A simplified, illustrative version of the same principle behind every production computation · computed right in your browser via SHA 256. Pick a scenario, then pick it again: the hash won’t change. That’s determinism, not coincidence · the reproducibility a model risk reviewer can test.
Runs entirely in your browser. Illustrates the principle · this is not the production risk kernel.
Six reasons we are taken on top of the existing stack
We don’t compete with Aladdin or Bloomberg for the role of primary system. We close what a large institution does not have: independent, deterministic, reproducible risk control · faster and on the same hardware.
Independence
A separate engine and separate mathematics from your front office · a genuine second opinion on model risk, not the same computation under a different name.
Deterministic engine
Tail metrics in milliseconds, without brute force Monte Carlo path enumeration · faster on the same hardware you cannot expand.
Audit grade trail
Hash pinned replay, reproducible, with every step audited · what model risk review can test (SR 26-2, FRTB, Basel).
Provable on your data
Computed on your book; you reconcile against Aladdin/Bloomberg and realized dynamics. Trust is built on your numbers, not our promises.
Embeds, does not replace
Runs alongside Aladdin, Bloomberg, Murex, MSCI as an independent control. Zero deployment, no integration projects.
Continuous control
Not a one time report, but ongoing recomputation with early warning on divergence from the front office and building risks.
Where we sit relative to what the largest institution already runs
We are not a third enterprise system in the row. We are an independent deterministic function on top of everything already deployed.
| Criterion | Your enterprise stack Aladdin · Bloomberg · SAS |
Internal validation own effort |
Big Four consulting audit / opinion |
SAA |
|---|---|---|---|---|
| Task | Primary risk computation | Self check | Opinion / audit | Independent deterministic challenger |
| Independence | One engine with the front office | Often not separated from the front office | External, but not reproducible | Separate engine and mathematics |
| Reproducibility | Depends on versions and seeds | Varies | No | Hash pinned replay |
| Speed on tails | Monte Carlo, does not scale | not applicable | not applicable | Deterministic, milliseconds |
| Deployment | Deployed, team | 100+ FTE staff | months of project work | no core replacement, on your data |
We don’t replace Aladdin or Bloomberg. We are an independent deterministic second opinion on top of them: what the regulator increasingly requires as a separate function, and what you cannot get by computing on the same engine or buying more GPUs.
The largest institutions on both sides of the market
You’ll recognize yourself if
- A large book where the tail is computed with Monte Carlo
- The regulator requires independent, reproducible validation
- Your validation team grows slower than the model count
- You need a second view of risk, independent of the front office
Large asset managers and hedge funds (buy side)
The buy side is the first to hit the compute shortage: TSMC’s advanced CoWoS packaging is sold out at least through 2027 (Silicon Analysts), and demand has tripled in two years · you cannot scale Monte Carlo by buying GPUs. Meanwhile a single Aladdin seat runs from roughly $1M a year (SmarterWay.AI) · and that’s the primary computation, not an independent check.
G SIBs and large banks (trading book)
Model counts are rising across every bank surveyed (joint Risk.net and Moody’s survey, 2026 · Moody’s), while validation is understaffed: even at banks with over $250B in assets the average shortfall is 18 positions against 115 already hired (RMA). SR 26-2 and FRTB IMA and SR 26-2 support an independent, reproducible challenger function · the lane we provide without adding validation headcount.
Pension and sovereign wealth funds
Large, long horizon portfolios are tail sensitive, and risk is often computed by external managers. An independent deterministic challenger gives the asset owner a second, reproducible view of VaR/CVaR and concentration · not dependent on the manager’s engine or assumptions.
Insurers and central banks
Insurers’ investment arms and central banks’ supervisory functions operate in the same model risk logic: they need a reproducible computation, independent of the front office, with an audit trail that withstands external scrutiny. ARIN22 delivers a hash reproducible result on your data, embedding alongside your existing systems.
The largest institutions already have systems that cost millions. What they don’t have is independent, deterministic, reproducible control that computes faster on the same hardware and produces a trail for the regulator. That’s our lane. We set no upper limit · the larger the book, the greater the effect.
What you receive
- •Challenger VaR/CVaR reportan independent computation against your front office, with divergences and their analysis.
- •Model risk assessmentunder SR 26-2 / FRTB IMA · with a reproducible trail of every step.
- •BacktestKupiec, Christoffersen, Acerbi-Székely (ES), EVT fit, PIT (power aware).
- •Tail and stressCVaR 99 / 99.5 / 99.9, scenario catalogue, deterministic deep tail.
- •Concentration mapTop 1 / 2 / 3, HHI, diversification · by counterparty and position.
- •Audit trail and monitoringhash reproducible, receipts, signals on divergence from the front office.
Institutional engagement · on request
Cost is determined by book size, number of models and depth of validation; agreed individually after a pilot on your data.
An individual quote after the pilot · scaled to your book and validation scope, not a fixed price list.
Early access program
Institutions onboarding at the current stage get preferential terms for the first year. Condition · consent to use anonymized results as a reference.
The essentials, briefly
Do you replace Aladdin or Bloomberg?
No. We are an independent deterministic challenger on top of them. Zero deployment, running alongside your existing stack.
How are you more independent than an internal team?
A separate engine and separate mathematics from your front office; the computation is reproducible hash reproducible by hash · that is the independence the regulator expects.
How does this help with the GPU shortage?
The deterministic kernel computes tail metrics in milliseconds on the same hardware where Monte Carlo takes orders of magnitude longer. More computation without buying more capacity.
How is data security handled?
Encryption in transit and at rest · NDA before any data is received · no resale of data · deletion on request.
How long does a result take?
From book handover to a validation packet · weeks. The exact timeline is fixed after scoping the data.
Independent validation of credit models · SR 26-2, IFRS 9, Basel III
The same independent deterministic check, but for credit risk: PD/LGD/EAD, IFRS 9, Basel III (RWA, capital, LCR/NSFR). For banks · under Fed/OCC/FDIC SR 26-2 (April 2026, risk based, supervising banks over $30B, superseding SR 11-7) and CECL provisioning. An external challenger closes the independent validation expectation without adding hundreds of FTE.
- Independent validation of credit models · under SR 26-2 / CECL and Basel III
- IFRS 9 (stages 1/2/3, lifetime PD), RWA, capital, LCR/NSFR
- PD/LGD/EAD, scoring, migration matrices, reproducible trail
London Whale · $6.2B on a single unvalidated model
JPMorgan took $6.2B in losses and paid $920M in fines over one unvalidated VaR model. A textbook argument for independent, reproducible validation · exactly what the regulator now requires as a separate function.
Tricolor · a collapse that hit every counterparty
Subprime auto lender Tricolor filed for bankruptcy on 10 September 2025 after warehouse lenders discovered double pledging; JPMorgan wrote off $170M (CNN Business). Independent collateral and concentration checks catch exactly these patterns before they materialize.
Honestly: the model ranks risk (AUC ≈ 0.69), but the macro overlay is directional, not a calibrated figure. The realized backtest is still accumulating its sample.
Discuss a credit model validation •What the two usual paths cost
Buy off the shelf
This is the primary risk computation, not an independent check. Sources: SmarterWay.AI, ZipRecruiter.
Build validation internal
A full function is years and hundreds of FTE. Sources: RMA, ZipRecruiter.
Buy off the shelf (Aladdin from $1M/yr) or build independent validation internal (hundreds of FTE) · both are costlier and slower than an external deterministic challenger on top of the stack you already run. We offer a third path: independence and reproducibility without expanding your GPU fleet or headcount.
Validate your risk independently · on your data
We’ll discuss a pilot: a challenger computation against your front office and the contents of the validation packet for your book.
Request a pilot