Explaining GBM pricing models to underwriters with SHAP

Gradient-boosted models nearly always beat a GLM on holdout deviance. The hard part is getting the people who own the rates to accept them. An underwriter won’t sign off a factor they can’t see, and they shouldn’t have to.

SHAP (SHapley Additive exPlanations) is the most practical fix I’ve found. Here’s how I use it and where to be careful.

What SHAP gives you

For each policy, SHAP splits the model’s prediction into a baseline plus one contribution per feature. For a Poisson GBM with a log link, contributions are additive on the log scale, so exponentiating them gives multiplicative effects. That is the same language as GLM relativities, which is the main reason SHAP works well for pricing audiences.

import shap, numpy as np

explainer = shap.TreeExplainer(gbm)          # LightGBM / XGBoost
sv = explainer.shap_values(X_holdout)        # log-scale contributions

# Implied multiplicative relativity by feature value
rel = np.exp(sv[:, X_holdout.columns.get_loc("sum_insured_band")])

Three views I show underwriters

1. Dependence plots per rating factor. The GBM’s implied curve set against the current GLM relativity. Where the two differ is where the discussion should be.
2. Interaction checks. Colour the dependence plot by a second feature (e.g. construction type) to show whether an effect really varies by segment.
3. Single-risk waterfalls. For referrals, show exactly why this risk is priced where it is.

Where it misleads

• Correlated features share credit arbitrarily. If sum insured and number of rooms move together, SHAP may split the effect between them in a way that has no causal meaning.
• Thin segments. In a small HNW book, a sharp SHAP curve at the extremes can come from a handful of claims. Always show exposure alongside the curve.
• Log vs response scale. Mixing them up in a slide is the fastest way to lose credibility.

How I use it in practice

I treat the GBM plus SHAP as a discovery tool. The signals that survive challenge are promoted into the GLM, which remains the rating engine of record. The GBM finds the structure, and the GLM keeps it explainable and governable.