Automobile Insurance Claims Risk Modeling

The business goal is to estimate the expected number of claims per policy per year so an insurer can price risk fairly — charging more for high-risk drivers and less for low-risk ones. The target is `ClaimNb` normalized by `Exposure` (claims per year).

The project is deliberately comparative: models hand-built from scratch sit next to library references and a classical statistical model, scored on the same split with the same metrics.

Claim data is heavily zero-inflated — the vast majority of policies never file a claim — so the work treats metric choice and distributional assumptions as central rather than incidental.

The dataset covers ~678,000 policies with 11 driver and vehicle features: driver age, bonus-malus, vehicle power/age/brand/fuel, region, population density, and exposure.

Claims are rare and over-dispersed, so plain accuracy is meaningless and squared error barely separates models.

Raw counts have to be normalized by exposure to model a comparable per-year claim rate.

It is not obvious up front whether modern ML actually beats classical actuarial regression on this kind of low-signal data.

A three-person BSc exam project at IT University of Copenhagen. The points below are my own contributions; teammates owned the from-scratch decision tree, the Random Forest, and parts of the final evaluation.

Built the feed-forward neural network from scratch in NumPy — including the custom loss functions, optimizers, batch iterator, and training loop.

Built the PyTorch reference MLP used to sanity-check the from-scratch network.

Implemented PCA from scratch and ran the dimensionality and clustering analysis of the policy data.

Led the data cleaning, preprocessing, and exploratory analysis pipeline, and contributed to the Negative Binomial GLM modeling.

Settled on log claim-rate — log of (ClaimNb / Exposure) — as the regression target to handle exposure and heavy skew.

Modeled the rare-event structure explicitly: a Negative Binomial GLM for over-dispersed counts, with exposure carried as a GLM offset.

Validated the from-scratch models by matching their test metrics against the scikit-learn and PyTorch equivalents.

Also framed a binary claim / no-claim task evaluated with ROC AUC, given how strongly zero-claim policies dominate the data.

The Negative Binomial GLM gave the best test RMSE (3.03) and an interpretable Poisson deviance of 1.88, edging out both the trees and the neural network.

The from-scratch decision tree matched its scikit-learn reference almost exactly (RMSE 3.04 vs 3.03), confirming the manual implementation was correct.

On the binary claim / no-claim task, a Random Forest reached ROC AUC 0.65.

Headline finding: on noisy, zero-inflated claim data the simple statistical model matched or beat far more complex ones — added complexity did not buy accuracy.