Simudyne

The models are the product in a literal sense, since agent based simulation constructs populations of individual actors with behavioural rules and lets system level outcomes emerge from their interactions, which is what allows the platform to run millions of scenarios and trace shock propagation paths. Its bank partner describes the combination as agent based modelling and artificial intelligence technology enabling rapid simulation at cloud scale.

Held at B because agent based modelling is a computational simulation discipline that long predates modern machine learning and does not depend on it, so stripping contemporary techniques leaves the simulation engine and its behavioural library intact, which is the same honest position recorded for Renew Risk.

The product exists so that people can rehearse rather than act, which is an unusual and genuinely sound position on this axis. Clients test drive decisions and, in the company's own phrase, fail fast without consequences, running millions of scenarios in a virtual environment before committing capital or changing a policy in the real one. Nothing is decided or executed by the system.

The qualification is downstream: simulation output is used to optimise execution algorithms and to set fraud detection thresholds, so conclusions drawn in the simulated environment do end up governing automated behaviour in the live one, and nothing describes what validation sits between the two.

A validation methodology is published and specified, with simulators validated through a six step process ensuring the model design and parameters reproduce statistical and behavioural dynamics matching real world observations, which is the correct standard for simulation because a model that does not reproduce known behaviour cannot be trusted on unknown behaviour.

Supporting it are six doctorates on a thirty person team, a published technical guide setting out the method, a public sandbox allowing inspection, and an argument grounded in academic literature on systemic risk. The company also states its case against the incumbent approach rather than merely asserting superiority, contending that stress testing fails to account for crisis dynamics, feedback, crowded trades, asymmetric information and liquidity shortages.

A globally systemic bank both led the Series A and published an unusually detailed account of using the technology itself, describing several use cases across credit risk, market risk and operational risk to model default contagion where a shock spreads between institutions or across geographies, with stress testing named as the next application, and characterising the platform as supplementing its existing scenario planning toolkit across multiple businesses and geographies.

Banks rarely describe a supplier's role in their risk stack that precisely. Revenue grew 600 percent year on year in 2018 on the addition of global banking clients, and the team of 30 included six doctorates across market simulation, fraud detection and risk management. The round also drew two venture firms known for backing well known consumer and banking names.

The architecture limits the usual concern by construction, since what the company supplies is a simulation engine and a library of behavioural functions deployed for the institution to run, rather than a service that ingests customer data into a shared model. Value comes from the method rather than from accumulated cross client data, so one bank's contagion analysis does not improve the engine another bank receives in the way an extraction or screening platform would. Held at B because no explicit boundary statement was located and calibration work necessarily touches client data.

Structurally favourable because of what simulation is. The platform replicates networks of customers, merchants and banks as modelled populations rather than processing records about identifiable people, so the fraud work tests behaviour that has not occurred rather than analysing behaviour that has, and no consumer is the subject of any assessment. Calibration to real world statistical dynamics implies access to institutional data, and that occurs inside the bank's own environment. Held at B because no data processing terms, retention schedule or subprocessor list was located.

No attestation, certification, trust centre or enumerated framework was located. A globally systemic bank has deployed the technology across three risk disciplines and invested in the company, which means supplier assessment has been passed at the most demanding standard available, and none of the resulting assurance is published for other institutions to read.

No supervisor, statute or instrument is named, and the omission is conspicuous because the flagship use case is a regulatory exercise. Stress testing exists because supervisors require it, with prescribed scenarios, submission timetables and model governance expectations in every major jurisdiction, and a platform whose stated purpose is improving on conventional stress testing operates directly inside that framework. The company argues the methodology point at length and never names the regime the methodology serves.

No individual is assessed and the adapted exposure is that a simulation embeds its builders' assumptions about how people and institutions behave. Agent based models are defined by the behavioural rules assigned to agents, and the company is explicit that agents act on idiosyncratic and potentially less than optimal rules, which is analytically honest and also means someone chose those rules.

Where the output sets fraud detection thresholds or informs capital adequacy, those chosen behaviours shape decisions affecting real customers, and a threshold tuned against simulated fraudsters adapting to detection may perform differently against real ones. No sensitivity analysis across behavioural assumptions was located.

No guarantee, indemnity or falsifiable commitment was located. The published validation process gives an institution a documented basis for assessing whether a simulator behaves correctly, which is more than most vendors provide and is what a model risk function would test against.

Nothing describes what happens when a simulation proves wrong in a way that mattered, which for this product means a stress test that understated a contagion path or a fraud threshold tuned against the wrong adversary, and no notification or correction process appears.

No external dependency is disclosed. The simulation engine and behavioural library are the company's own, and the method itself is published in technical detail, which is a form of transparency most vendors do not offer even though it is about technique rather than about suppliers.

What is absent is everything else: no cloud provider is named despite cloud scale being central to the proposition, no market or reference data source is identified for calibration, and no subprocessor list appears.

The platform is described as enterprise ready, scaling on modern cloud infrastructure and offering a library of functions for financial applications, which speaks to deployability rather than to connection. No named integration appears on either side: no risk system, market data source, core banking platform or model governance tool is identified, and no developer documentation was located, although a public sandbox environment demonstrates the capability.

Cloud infrastructure is referenced as enabling massive scale and no provider, region selection, residency commitment or on premise option is published. The question matters less than for platforms holding customer records, since simulation runs on modelled populations, and a bank calibrating models against its own portfolio data would still expect a stated position on where that computation occurs.

No pricing, packaging or basis of charge was located. The platform is described as enterprise ready software with a library of financial functions, which suggests a licence model, and nothing indicates whether charge scales with users, simulations run, compute consumed or modules taken. A public sandbox demonstration exists, which reduces evaluation cost without addressing price.

Functional coverage is unusually broad for one technique because the method generalises: stress testing, systemic contagion, market execution optimisation, financial crime and anti money laundering, fraud, credit, market and operational risk, and corporate loan prepayment analysis using a competing risks survival framework all run on the same engine. The entities modelled span banks, asset managers, funds and market infrastructures, which is the whole system rather than one firm. What is evidenced on the buying side is narrower, being large banks specifically, with stated ambitions to enter other sectors.