Automatic Generalized Linear Models: Turbo‑charging Interpretable Models for Financial Services

In regulated financial environments, accuracy alone is not enough. Pricing and risk teams need models that are explainable, auditable, and easy to deploy, while still delivering strong predictive performance. 

Generalized Linear Models (GLM) remain the industry standard because they are transparent, well understood, and trusted by regulators. But modern datasets rarely behave linearly. Important effects are often non-linear, segments behave differently, and interactions matter. Capturing this complexity with traditional GLMs requires extensive manual effort and slow iteration cycles. 

Earnix Automatic GLM (AGLM) was built to remove that bottleneck. It preserves the GLM structure that regulators expect, while automating the most time-consuming parts of feature engineering. The result is a model that fits naturally into governance and deployment workflows, while delivering much of the predictive lift associated with modern machine learning. 

Why GLM Still Matters 

Before diving into automation, it’s worth remembering why GLMs are so entrenched in finance: 

• Widespread acceptance and regulatory familiarity.

  GLMs and their extensions have been used for decades; many practitioners regard them as the status quo for modelling. Insurance rate‑filing systems, tariff modelling, and credit‑scoring engines are built around GLM outputs. 

• Interpretability and inference.

  A GLM expresses the expected response as a linear combination of predictors after a link function. This structure enables one‑to‑one mapping from features to outcomes and supports inference: confidence intervals, significance tests, and hypothesis checking. Risk professionals can explain the contribution of each factor to stakeholders without resorting to surrogate models or SHAP plots. 

• Efficiency and control.

  GLMs are computationally light and easy to implement. They allow modellers to decide which variables to include, choose link functions and transformation terms, and control the model structure. In regulated industries, where justification and audit trails are essential, such control is invaluable. 

What Automatic GLM Changes 

AGLM preserves the value of GLMs - transparent structure and defensible outputs - while automating three tasks that typically demand the most modeling time and expertise: 

1. Discovering important interactions 

   In many financial datasets, the most important signals are not single variables—they’re combinations (for example, how vehicle type behaves differently across driver ages, or how customer tenure interacts with product mix). Exhaustively searching interactions doesn’t scale, while relying only on intuition risks missing critical structure. 

   Automatic GLM uses a data-driven approach to propose a short list of high-impact interactions, so you get meaningful lift without exploding complexity.  

2. Capturing nonlinear effects, without hand-crafted binning

   Traditional GLMs often require manual binning (or careful spline design) to handle nonlinear relationships. That’s slow, subjective, and hard to standardize across teams. 
   Automatic GLM automatically learns a compact, interpretable representation of nonlinear effects by starting with granular bins and then merging where the data supports it, so the model can flex where it needs to and can stay simple where it doesn’t.  

3. “Smart grouping” for high-cardinality categorical features

   Categorical variables with many levels are a classic pain point: leaving them as-is can generate noisy, unstable estimates; manually grouping them is time-consuming and often inconsistent.

   Automatic GLM clusters categories with similar impact, ensuring the final model remains readable, stable, and defensible, even when categorical features are large.  

Crucially, after these transformations and selections are made, the workflow produces a final model that is still a GLM—with coefficients you can interpret, validate, and explain. And because it’s a native Earnix GLM, it remains fully editable and customizable in the platform: you can review the structure, adjust features and settings, and refine the model just like any other GLM before moving it into production. 

Think of it as a GLM-ready output, without the GLM-heavy workflow. AGLM reduces manual binning and interaction hunting, speeds up iteration, and helps teams standardize model development—while keeping the model defensible in governance. 

The Application of AGLM in Financial Services 

AGLM is designed for pricing and risk teams that need models that are both explainable and deployment-ready, specifically models that can be converted seamlessly and accurately into a rating-table structure within Earnix workflows. 

• Insurance pricing & reserving: capture nonlinear rating effects and key interactions while preserving tariff transparency and clean translation to rating tables. 

• Credit scoring & risk modeling: improve predictive power while maintaining governance-friendly interpretability and a structure that fits regulated decisioning processes. 

• Operational deployment in Earnix: move from model development to production faster by generating a model output that aligns naturally with how Earnix customers operationalize rates—supporting consistent implementation, review, and change management. 

Performance Benchmark (MTPL claim frequency) 

To quantify how AGLM compares with both classic interpretable methods and modern machine learning, we ran a benchmark on the French and Belgian MTPL claim frequency datasets (motor third-party liability). We trained each model to predict claim counts using a Poisson setup with exposure as an offset and evaluated performance on a 20% hold-out set. 

We report Explained Poisson Deviance (higher is better), which captures how much improvement the model provides over a null baseline—often treated as an R²-like measure for count models.