Lumnion

The modelling layer is the product and it is not decorative: gradient boosted trees, random forests, decision trees and both generalised linear and generalised additive models, fitted on the carrier's portfolio, compared side by side, with micro segment detection and a real time advisory module over the portfolio.

Held at B on the pocket's own residue rule, which now holds four times: a deterministic rate engine, a data preparation module and an external data module remain underneath if the models are removed, exactly as with the three larger pricing vendors here. The single exception in this pocket is the one vendor that never had a rating engine at all, which is the cleanest possible statement of where the line sits.

Oversight is built into the tool rather than asserted around it: model comparison on a common dataset, statistical results for every factor, factor and interaction advice, impact analysis and scenario simulation before a price is committed, and audit traceable reporting afterwards. The actuary drives every step and the product is designed for inspection.

Held off A because the rate engine is sold on pushing a commercial price decision into the market instantly and nothing names an approval gate, threshold or enforcement mechanism standing between a simulated scenario and a live tariff.

Substantial named machinery aimed squarely at letting the institution validate the model itself, which is the route to a high grade on this axis: side by side model comparison on a common dataset, statistical results for all factors and models, factor and interaction advice, variable grouping, automated reporting and audit traceable output. Held off A on a distinction this pocket needs and the market blurs.

The central claim is a proprietary methodology that CONVERTS the output of opaque models into base prices and coefficients, and no published methodology, fidelity measure or validation of that conversion exists. Converting a black box into coefficients after the fact is not the same thing as fitting an interpretable model in the first place, and the difference is whether the conversion itself has been validated. The pocket anchor generates interpretable models directly; this vendor explains opaque ones afterwards, and both are sold under the same word.

Seventeen named customers published on the company's own site, including three global insurance groups and a national insurance sector information and monitoring body, which for a company of roughly eighteen people is a substantial installed base. Held at B because not one outcome is attached to any of them: no case study, no quantified result, no executive on the record. The largest names carry the least evidence, which is the pattern this index keeps finding, and the practical effect is that a buyer can verify who bought it and nothing about what it did.

Nothing published on whether a carrier's portfolio, claims or quote data informs anything beyond that carrier's own models. The specific gap is in the external data module, which is sold as supplying statistically aggregated information from trusted sources and names not one of them. For a pricing input that will enter a filed or published tariff, an unnamed data source is a material silence, and it is the mirror image of this vendor's unusually complete disclosure on the algorithm side.

Rare shape in this index: the privacy posture is evidenced by an independently audited privacy information management certification with the certificate published, rather than by a policy page asserting good intentions. The operating entity is German and therefore inside the European regime by construction.

Held off A because there is no data processing agreement, no subprocessor list, no residency statement, and nothing describing what the vendor may do with the policy, claims and quote data a carrier loads into the modelling environment.

The best credential presentation found anywhere in this sweep, and it is worth using as the reference example. Four current certifications, each with the edition year stated, the auditing body named, and the certificate itself published for download in one click: information security management to the 2022 edition, privacy information management, information technology service management, and quality management.

Nothing that is not a certification is mixed into the row, no statute or framework is padded in, and no analyst badge is smuggled alongside. That is the exact inverse of the badge row failures catalogued elsewhere here. The privacy management certification is genuinely uncommon in this index.

Banked check, and it is the right one for a presentation this strong: whether the certification body is itself accredited by a recognised national accreditation body, since a certificate is only worth the accreditation behind the certifier. Also absent: any service organisation control report, a trust portal, penetration testing disclosure and any description of the controls themselves.

Unlicensed and unsupervised, which is the norm for this pocket. One customer is a national insurance sector information and monitoring body, a statutory institution, and that is a notable reference rather than any form of standing. No supervisory examination, no filing approval and no sandbox participation published.

No fairness testing, no protected class position, no governance statement. And this vendor states the mechanism more plainly than anyone else in the pocket has: alongside risk modelling it sells commercial optimisation and behavioural pricing, described as taking pricing to a personal level through behavioural optimisation. That is individualised price optimisation against customer behaviour, named as a product capability, with nothing beside it on who it disadvantages. Fourth instance of the pocket pattern, which is now four vendors selling price optimisation with silence against two selling risk modelling that disclose bias testing.

No recourse route, no allocation of responsibility, nothing beyond standard terms of use. The policyholder priced by a behaviourally optimised tariff has no visibility of the model, the coefficients derived from it, or the external data enriching their record, and no route to any of the three.

Every algorithm family the platform can run is named on the company's own site, the models are fitted on the customer's own portfolio, and no third party model service sits in the path, so the supply chain is essentially fully described.

That is better disclosure than most of this index achieves and it costs the vendor nothing, which is the point worth carrying: for a classical machine learning vendor the algorithms are public and the training data belongs to the buyer, so the question is cheap to answer. The vendors that cannot answer it are the ones that bought a language model and will not say whose. Held off A because the external data module's sources are unnamed, so one input to the price remains undisclosed.

Positioned explicitly as connecting to any core system, delivered through cloud interfaces with an integrated rule and rating engine so a price change reaches the market without a core release, which is the same strangler pattern the other modern vendors in this pocket sell. Listed on the Amazon marketplace as a subscription service. A named partnership with a core insurance and banking platform vendor that is itself in this index is a real integration signal rather than a logo. Held off A because no named connector, protocol or reference implementation is described.

Cloud service on Amazon infrastructure with a marketplace listing and interfaces for integration, and a certified service management system covering support, with a published commitment to respond within twenty four hours. Nothing published on region, residency, tenancy or hosting options, which is a live gap for a German entity selling to German and Turkish carriers whose supervisors take different positions on where policy data may sit.

No price, tier or unit published on the company's own site. Two partial signals worth recording: free modules are offered as an entry point, and the marketplace listing states a recurring monthly subscription billed through the cloud provider, though no figure appears in the public listing. A buyer can start without talking to sales and still cannot budget.

Non life and health carriers, sold in two core markets (Turkey and Germany) with global groups' local entities among the named customers. Seventeen named insurers spanning motor, property, health and life pension shapes, plus a national sector data body.

Held off A because the buyer type is narrow by design (the carrier actuarial and pricing function, not brokers, managing general agents or reinsurers) and the geographic base is two markets rather than a genuinely global one, on a headcount of roughly eighteen.