---
title: "A Flat Large-Claim Loading Undercharges Young Drivers"
pubDate: 2026-11-18T02:00:00.000Z
tags:
  - Statistics
  - Insurance
  - Math
math: true
image: /images/blog/large-claim-loading/featured.png
---

import PricingFigure from '../../components/PricingFigure.astro';

## TL;DR
In the French motor portfolio freMTPL2, the largest 1% of claims carry 38% of the cost, enough to make one cross-validation fold look almost twice as expensive as another. So every model is fitted to claims capped at their 99% quantile, and the excess, a third of the cost, comes back as one loading $\ell = 1.51$ on every premium. That is fair only if every claim carries the same expected excess. Tested per rating factor, it holds for all but driver age: claims of drivers up to 21 are large twice as often, and the three largest excesses are theirs, so the flat loading charges them about 78% of their expected cost. And $\ell$ itself, with a 95% interval from 1.28 to 1.80, is the most uncertain number in the tariff.

*Second of four posts on pricing freMTPL2, after [the data](/post/fremtpl2-head-block) and before [the model comparison](/post/tweedie-glm-vs-xgboost).*

## Why cap

### The problem is variance, not bias

The models are trained on a random 80% of the cleaned portfolio, 534,883 policies with 20,879 claims. The cap $c$ is the 99% quantile of their claim amounts, 18,000:

<PricingFigure name="split" />

*Training claims split at the cap: 1% of the claims lie above it, and their part above it is a third of the cost.*

Every quantity a pricing model estimates is a mean of claim amounts over some set of policies: a fold, a segment, a level of a rating factor. Where a handful of claims carry a third of the cost, that mean is set by those few amounts, and moving one of them to another set moves the estimate. Claim frequency is the same in all five cross-validation folds, between 0.073 and 0.075; the pure premium is not:

<PricingFigure name="folds" />

*Pure premium per policy-year in each fold and the test set, uncapped and capped.*

Uncapped, a model's score on a fold would depend more on which fold got the largest claims than on the model. The same goes for relativities: in the youngest driver-age band, three claims, the three largest in the portfolio, make up 71% of the raw cost. Capped, the band still costs 2.1 times the next one, an effect all of its claims inform.

### Cap each claim, load the excess

Each claim is capped on its own, $Z_{ik}^{(c)} = \min(Z_{ik}, c)$, before it is added to the policy total $S_i^{(c)}$. Every model is fitted to the capped pure premium and predicts $\hat\mu^{(c)}(x)$, and the premium charged is

$$
\hat\mu(x) = \ell\, \hat\mu^{(c)}(x), \qquad \ell = \frac{\sum_{i \in \mathcal I_{\mathrm{train}}} S_i}{\sum_{i \in \mathcal I_{\mathrm{train}}} S_i^{(c)}} = 1.508.
$$

If the capped model is balanced, its premiums adding up to the capped training cost, then the loaded premiums add up to the full training cost. So the cap moves cost between policies, from those that happened to have a large claim to everyone, but not out of the tariff. The question is whether it moves it fairly.

## What a flat loading assumes

Split each claim at the cap, $Z = Z^{(c)} + (Z - c)_+$, and write $m^{(c)}(x) = \mathbb E[Z^{(c)} \mid x]$ for the mean capped claim and

$$
\kappa(x) = \mathbb E[(Z - c)_+ \mid x] = \rho(x)\, \mathbb E[Z - c \mid Z > c, x], \qquad \rho(x) = \mathbb P(Z > c \mid x),
$$

for the expected excess per claim: how often a claim is large, times by how much. With claim frequency $\lambda(x)$, assume

- **(L1)** every claim carries the same expected excess, $\kappa(x) = \kappa$: a large claim is an ordinary claim that turned expensive, which can happen to anyone;
- **(L2)** the mean capped claim doesn't depend on the features, $m^{(c)}(x) = m^{(c)}$.

Then

$$
\begin{aligned}
\mu(x) &= \lambda(x)\,\big(m^{(c)}(x) + \kappa(x)\big) && \\
&= \lambda(x)\,\big(m^{(c)} + \kappa\big) && \text{(L1), (L2)} \\
&= \Big(1 + \frac{\kappa}{m^{(c)}}\Big)\, \lambda(x)\, m^{(c)} = \ell\, \mu^{(c)}(x), &&
\end{aligned}
$$

and the ratio of full to capped cost estimates exactly this $\ell$. (L2) is mild: capped claim size is close to flat across the rating factors. (L1) is the strong one, and it can be tested on the 210 training claims above the cap.

## Testing (L1)

**How often a claim is large.** Per rating factor, a [χ² test of independence](https://en.wikipedia.org/wiki/Chi-squared_test) asks whether the share of large claims is the same at every level. Five of seven factors show nothing at the 5% level, from vehicle power ($p$ = 0.21) to brand ($p$ = 0.93). Driver age does:

<PricingFigure name="largeshare" />

*Share of training claims above the cap per driver-age band, with exact 95% intervals; dotted, the share over all claims.*

Claims of policyholders aged up to 21 are large 2.3% of the time, 17 of 753, against 1.0% for everyone else. With so few large claims per band the χ² approximation is rough, so this one comparison gets [Fisher's exact test](https://en.wikipedia.org/wiki/Fisher%27s_exact_test): $p$ = 0.002.

**How large a large claim is.** The typical excess is the same for young and older drivers, medians of 18,337 and 19,622, and a rank-based [Mann–Whitney test](https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test) gives $p$ = 0.61. The top is not: the three largest excesses are all young drivers', although only 17 of the 210 large claims are. If size ignored age, that would happen with probability $\binom{17}{3} / \binom{210}{3} \approx 0.04\%$. The group was chosen after seeing those claims, so read this as a measure of surprise, not a test.

**What it costs.** For age, (L1) fails in both parts. If young drivers' claims are large at their own rate $\hat\rho$ and exceed the cap by the pooled mean excess, the flat $\ell$ charges them

$$
\frac{\ell\, m^{(c)}_{\text{young}}}{m^{(c)}_{\text{young}} + \kappa_{\text{young}}} \approx 78\%
$$

of their expected cost, and everyone else 101%. Since the pooled excess ignores their heavier tail, the shortfall of 22% is a lower bound.

Dropping the cap wouldn't fix it: the young band's own ratio of full to capped cost is 6.0, set by three amounts, the variance problem the cap exists for. What should change is how the excess is spread. Give it its own frequency,

$$
\hat\mu(x) = \lambda(x)\,\big(m^{(c)}(x) + \hat\rho_{g(x)}\, \bar\kappa\big),
$$

with the share of large claims $\hat\rho_g$ per age group $g$ and one pooled mean excess $\bar\kappa$. The young-driver loading then rests on how often their claims are large, which 753 claims can estimate, not on how large three of them were. A [generalised Pareto](https://en.wikipedia.org/wiki/Generalized_Pareto_distribution) tail by age would be the next step.

## The most uncertain number in the tariff

The excess sits in 210 claims, so $\ell$ is noisy. Resampling the training policies gives its bootstrap distribution:

<PricingFigure name="loading" />

*The large-claim loading with its 95% bootstrap interval, the value without the single largest training excess, and the value the test set would have needed.*

The single largest excess is 24% of all excess cost. The test set had only 30 policies with a claim above the cap and would have needed $\ell = 1.12$, so the loaded premium predicts 1.37 times its full cost, while the capped premium is within 2% of its capped cost. That isn't a model failure: a year with a few fewer large claims looks exactly like this. It does mean that the factor multiplying every premium is known only to within about ±20%, while the [model comparison that follows](/post/tweedie-glm-vs-xgboost) fights over fractions of a percentage point. Several years of claims, or a tail model above the cap, would improve the premium more than a better model of the capped part could.

A flat factor is a modelling assumption like any other. Written down, it turned into two testable statements, and the test found who pays for it.
