Halley shows banks and insurers what their machine learning models have really learned.

Models now decide who gets a loan, what a policy costs, and which transactions get flagged as fraud. Model risk, compliance, and business teams are accountable for those decisions - but standard explainability tools reveal only a fraction of what a model is actually doing.

We find what they miss: complex multi-feature interactions the development team didn't know were there, and subgroups where the model is systematically wrong. Two individually harmless features in a credit model that combine to reject borrowers who would have repaid. A pricing model that systematically underprices its riskiest segment. A fraud model that has overfitted on last year's attacks and will miss next year's.

These behaviors are invisible to SHAP and attribution methods, and computationally prohibitive to find by brute force. Our proprietary technology is model-agnostic - it works on anything from XGBoost to deep neural networks - and has been validated in collaborations with MIT, Meta, and UCL.

We're looking for pilot partners. If you lead model risk, validation, data science, or actuarial teams at a bank or insurer, or know someone who does, get in touch: www.halleysystems.com/#contact

Halley Systems

Org chart

Teams

Offices

Halley Systems

Org chart

Teams

Offices

Related companies