Murex

Clearwater precedent, cleanly. MX.3 has been a complete cross asset trading, risk and post trade platform since 1986. Strip every neural network and the entire platform remains and continues to price, book, clear and settle, only more slowly and at higher hardware cost. That is the vendor's own stated value proposition: machine learning is presented as a way to bring computationally expensive models into mainstream usage without prohibitive infrastructure investment.

Acceleration, not the product. The Murex AI Research initiative launched July 2026 is described by the company as focused on long term exploration, so it earns nothing under the honest scoping rule. Recorded because the shape is genuinely different from the rest of the platform tier: peers add assistants beside a core system, whereas Murex has changed how a core computation is performed. That still does not make the model the product.

Nothing published on oversight of the machine learning pricing path. No tolerance band, no fallback to the exact model, no threshold at which a valuation is rejected or recomputed, no statement of who reviews a surrogate priced number before it reaches a book. Worth stating precisely because the exposure is unusual: the neural network approximates a model whose exact answer Murex already possesses, so a comparison control is architecturally available and is simply not described.

General product controls exist and are credited under security rather than here: role based access control, and a four or more eyes principle on sensitive data changes. Those govern human actions in the platform, not model outputs.

The Loxon shape, and one of the better instances in the index: a vendor that publishes primary model work rather than explainability language. Murex has published a white paper on derivatives pricing with neural networks describing the neural network architecture chosen and how it must be trained to reach precision and reliability sufficient for trading and risk use, its head of quantitative research is named on the record discussing model development and evaluation, and the AI Research programme publishes to arXiv and supports PhD work with named universities.

The structural point is stronger still and worth carrying: because the network is trained to replicate a model Murex already possesses in exact form, fidelity is measurable against a known answer rather than against a judgement. Held at B: no published fidelity or error bounds for the shipped models, no accuracy figures, no drift monitoring, no versioning policy and no external validation. Methodology published, results not.

A deep named client set with a standing case study library: DZ Bank, Rabobank and Rabobank Brazil, Julius Baer, Daiwa Securities moving to MXSaaS on Amazon Web Services, Banorte, Aldermore and Momentum Metropolitan, plus go live announcements for Helaba, Abanca Portugal and a precious metals group.

Independent analyst recognition is heavy and current: Chartis category leader for Enterprise Market Risk 2026, fifth in Chartis Quantitative Analytics50 and eighth among technology vendors in BuySideRisk50, first in the IBSi Sales League Table for an eighth consecutive year, and repeat Risk.net awards. Held at B rather than A on one point: not one named institution has a quantified outcome attached, and none of the client stories is about the machine learning capability at all. The only hard numbers near the AI, a 7x performance improvement and 4x energy reduction, come from Nvidia describing benchmarks run in Murex's own research lab, not a client deployment.

A distinctive architecture with no commitment attached to it. Murex states that it performs training on its own infrastructure and then distributes the trained models to many clients, where execution is limited to inference. Training inputs are described as high dimensional derivative contract data and term structure data such as yield curves and volatility surfaces.

Nothing states whether any of that originates from client books or is generated, whether one client's positions inform a model another client runs, or what is retained. A centrally trained model shipped to many clients is the third stewardship shape seen on this roster after bring your own model and build it on your own data, and like both of the others it is described as an engineering fact and never as a commitment.

A standard privacy policy and nothing addressed to the platform. No retention terms, no statement of what client data Murex personnel can access in the managed service, no processing commitments specific to MXSaaS or BPaaS where the vendor operates the environment rather than the client. The privacy question is materially larger under the managed models than the licence model, and the site does not distinguish them.

The inverse of the badge row problem this index usually documents: strong published controls, almost no badges. The security page describes named mechanisms rather than intentions, including identity management with fine grained role based authorisation, MFA, SSO and SAML, a four or more eyes principle with audit on sensitive data changes, encryption in transit and at rest, continuous code review and static analysis, and regular penetration testing with remediation before release.

The single credential is a SOC 2 Type 1 attestation on MXSaaS, and the vendor uses the correct noun, attestation rather than certification, and scopes it precisely to the SaaS offering rather than the company. That precision is rare here and is credited. Held at B: Type 1 tests design at a point in time rather than operating effectiveness over a period, no ISO 27001 appears anywhere, and there is no trust portal, report period or bridge letter against the standing reference bar.

A software licensor, unlicensed and unsupervised in its own name. Recorded because it is the cleanest live instance of the open question left by the bank service provider work: Murex both licenses MX.3 to institutions that run it themselves and operates MXSaaS and BPaaS, performing the processing on the client's behalf.

The settled rule is that supervisory authority attaches to performing the service rather than to licensing software, which would place the managed side inside scope and the licence side outside it. No examination, registration or supervisory relationship is claimed by the vendor anywhere, so nothing is credited. Queued rather than refused silently: if any supervisory examination of the managed services entity is publicly identifiable, this axis is reconsidered.

Nothing published. Recorded with the honest qualification that this axis has genuinely low salience for the shipped capability: a neural network approximating a derivative valuation has no protected class dimension and no individual subject, unlike credit decisioning or identity proofing where a C is a real gap.

The forward looking exposure is real though and is worth naming: the AI Research programme lists explainable, trustworthy and responsible AI as a research priority and intelligent automation across the capital markets value chain as another, so the company has stated an intention to move models toward decision support without publishing any governance position to carry it.

Sharp, and sharper than the platform tier norm because of where the output lands. A surrogate model that departs from the exact model it replicates produces a wrong valuation, and Murex's own material places XVA measures on a path that runs from pricing through accounting to capital requirements. A fidelity failure therefore propagates into traded prices, hedge decisions, reported profit and loss, audited accounts and regulatory capital.

Nothing published says who bears that, how a mispriced trade is identified after the fact, or whether the exact model is rerun as a check. The Abrigo test applies directly and fails: for a vendor whose output enters the financial statements, ask whether there is a SOC 1 and not only a SOC 2. Only a SOC 2 Type 1 exists.

A third route to B on this axis, alongside naming the provider you chose and handing the choice to the customer: build the model yourself and publish its architecture. The shipped pricing models are Murex's own, trained by Murex, and the company has published a white paper setting out the network architecture and the training approach, with a summary of that work independently described in trade coverage and its research arm publishing openly to arXiv.

That is more useful to a buyer assessing what sits underneath than a supplier name would be. Scoping stated explicitly rather than glossed: this covers the pricing stack only. The lower rungs are also present and earn nothing on their own, per the standing rule, namely AWS as a cloud, NVIDIA as hardware and Quant Network as a digital asset partner. No model provider is named for any generative capability, and none appears to be shipped. Queued check: the white paper itself has not been opened, so the scoping question of whether it describes the shipped stack or a research prototype is live, and a re grade should start there.

MX.3 is itself the system of record for trading, risk and post trade, so integration depth is structural rather than claimed. Published specifics rather than adjectives: a tiered service oriented architecture, a documented REST interface layer the vendor put in client hands at a two day hackathon, documented connectivity, and stated throughput envelopes including 1,000 imported trades per second at peak with stable end to end latency and full exposure and margin calculation across 20,000 collateral agreements in ten minutes. Publishing a performance envelope is the same class of evidence as naming a messaging standard or publishing a schema: it turns an integration claim into something a buyer can test against.

Four distinct delivery models, each separately named and documented: on premise licence, cloud, vendor managed MXSaaS, upgrade as a service, and a BPaaS managed offering, with AWS named as the cloud underneath at least one live client. A secure deployment guide is provided for both customer site and cloud installation, and the architecture page states the client access paths and the performance envelope rather than asserting scalability.

First deployment B of the platform tier and the reason is specificity, not option count. Held off A because the second half of this axis is unaddressed: no data residency statement, no regional hosting map and no sovereignty commitment, for a platform operating across 60 countries. One case study notes that a Brazilian implementation meets local requirements, which implies localisation capability without committing to anything.

No pricing published anywhere. Every route is a contact form. One piece of real commercial information survives and is recorded without moving the grade: the cloud based BPaaS managed service is described as pay as you go with minimal configuration and no additional infrastructure, and is positioned explicitly at smaller banks previously priced out of trade pricing and XVA management. That discloses a pricing model and a target buyer without disclosing a price, and the axis measures what a buyer can learn about cost.

About 300 client institutions, 60,000 daily users, 60 countries, stated by the vendor. Six named buyer segments each with its own page: banks, clearing houses, energy and commodities firms, investment managers, corporates and public agencies. MX.3 underpins a swap clearing platform carrying roughly 90 percent of OTC vanilla swap volume and processes over 35 million foreign exchange cash trades a day.

Named institutions across the case study library reach from global banks to national and regional ones: DZ Bank, Rabobank, Julius Baer, Daiwa Securities, Banorte, Aldermore, Helaba, Garanti BBVA, Abanca Portugal and Momentum Metropolitan, alongside a precious metals group. Among the broadest coverage in the index on any reading.