Throughout the financial services space, modelling teams face a persistent and increasingly uncomfortable problem: accuracy alone is no longer enough. Insurers and lenders are under pressure to extract more predictive power from growing datasets, yet they remain constrained by regulatory expectations that demand transparency, auditability and clear justification. This tension is most visible in pricing and risk functions, where models must be both technically robust and defensible to regulators.
Earnix, an InsurTech specialising in pricing and rating optimisation, has built its latest modelling capabilities around this exact challenge. The firm is targeting organisations that rely heavily on Generalised Linear Models (GLMs) but are struggling to keep pace with the complexity of modern data. While GLMs remain the industry standard, many insurers now find that traditional approaches are too slow and rigid to capture non-linear behaviour and shifting customer dynamics.
For decades, GLMs have been the default choice for regulated modelling. They are familiar, computationally efficient and, crucially, trusted by supervisors. Insurance rate filings, credit decision engines and reserving frameworks are all built around GLM outputs.
However, the simplicity that makes GLMs attractive from a governance perspective also limits their ability to reflect real-world patterns. Important interactions are often missed, and non-linear effects are flattened through manual assumptions.
Addressing these shortcomings using classic GLM techniques is possible, but rarely efficient. Actuaries and data scientists must manually bin variables, design spline structures and hypothesise interactions based on experience.
This process is time-consuming, subjective and difficult to standardise across teams and geographies. As a result, firms are often forced into a compromise between explainability and performance.
Earnix’s Automatic GLM (AGLM) is designed to reduce that compromise. Rather than replacing GLMs with opaque machine learning models, AGLM automates the most labour-intensive elements of GLM development while preserving a structure that remains regulator-friendly. The objective is to help teams uncover additional predictive lift without introducing governance risk.
AGLM focuses on three core areas. It automates the identification of high-impact interactions, allowing models to reflect how variables behave differently across segments without creating unmanageable complexity. It also captures non-linear effects without relying on manual binning, starting from granular groupings and merging them where supported by the data. Finally, it applies “smart grouping” to high-cardinality categorical variables, clustering similar categories to improve stability and interpretability.
Crucially, the output remains a fully transparent GLM. Coefficients can be reviewed, tested and adjusted before deployment, and models integrate directly into Earnix’s pricing and rating workflows. This enables smoother transitions from model development to operational use, particularly in environments where rating tables and formal approval processes are required.
Benchmark results from French and Belgian motor third-party liability datasets highlight the potential of this approach. In claim frequency modelling, AGLM delivers performance close to gradient-boosting techniques such as CatBoost, while consistently outperforming traditional interpretable methods like MARS. The results suggest that a significant portion of machine learning performance gains can be achieved within a governed modelling framework.
For insurers navigating increasing regulatory scrutiny and competitive pressure, the implications are significant. The long-standing trade-off between interpretability and accuracy is becoming less rigid. By automating feature engineering within established modelling practices, Earnix is positioning AGLM as a pragmatic path for insurers looking to modernise pricing and risk models without abandoning regulatory trust.
Read the full blog from Earnix here.
Copyright © 2026 FinTech Global
