Rishav Sharma Swiss Re

GitHub LinkedIn Medium Email CV

Swiss Re · 2023 → 2024 in review

Loss-ratio hotspots

Training a model to predict losses, then asking it which factors drove them.

Python scikit-learn SHAP Pandas

This was built inside Swiss Re, so the data and model aren't public. What follows is the idea and the outcome from my CV, nothing internal.

The problem

An insurance portfolio loses money in some places and not others. The loss ratio is claims paid divided by premiums earned; above 100%, the insurer pays out more than it takes in. The question was not "what will the loss be?" but "where, and why?"

The idea: explain the prediction

A model that predicts losses well has learned something about what drives them. SHAP is a way to ask it: for any one prediction, it splits the result into each input's contribution, fairly, even when inputs interact.

claim frequency policy tenure region lowers predicted loss raises it high value low value
Fig. 1 — Each dot is one policy. Its position shows how far one factor pushed that policy's predicted loss.

Read the picture row by row. Dots right of the centre line are policies where that factor raised the predicted loss; dots to the left, where it lowered it. Filled dots are high values of the factor. A row where filled dots sit far right says: high values of this factor drive losses up.

What came of it

A key-driver analysis on a US portfolio of agents and lawyers, with SHAP on top of the model, gave a 2.5% signal uplift over the base model. It surfaced drivers such as claim frequency, policy tenure and regional risk; the changes that followed brought loss ratios down by 5%.