French Motor Third-Party Liability Dataset

Overview

In this section, we prepare the French Motor Third-Party Liability (freMTPL2freq) dataset, a standard benchmark in actuarial science. Our goal is to transform the raw data into a clean, structured format suitable for both R and Python environments.

Data Preprocessing

The following transformations are applied to handle outliers, improve model stability, and align with actuarial pricing practices:

  • Target Variable: We define Frequency as the number of claims per unit of exposure. \[\text{Frequency} = \frac{\text{ClaimNb}}{\text{Exposure}} \]
  • Capping: VehAge (vehicle age) is capped at 25 years to mitigate the influence of extreme outliers (vintage cars) which follow different risk profiles.
  • Log-Transformation: Density (population density) is log-transformed to stabilize the variance and simplify its relationship with claim frequency.
  • Categorical Lumping: For VehBrand and Region, infrequent levels are collapsed into an “Other” category. This prevents overfitting in low-exposure segments.
Code 
# load dataset from CASdatasets
data(freMTPL2freq, package = "CASdatasets")

# preprocess dataset
df_all <- freMTPL2freq |>
  dplyr::mutate(
    Frequency  = ClaimNb / Exposure,
    VehAge = pmin(VehAge, 25),
    VehBrand = forcats::fct_lump(VehBrand, 5),
    LogDensity = log(Density),
    Region = forcats::fct_lump(Region, 6)
  ) |>
  dplyr::select(
    Frequency, Exposure, VehPower, VehAge,
    DrivAge, VehBrand, VehGas, LogDensity, Region
  ) |>
  dplyr::as_tibble()

Data Split for Hold-Out Validation

To ensure a robust evaluation of our models, we split the dataset into training and testing sets.

We store these datasets in Parquet format. Unlike standard CSV files, Parquet preserves schema information (e.g., categorical types) and provides high-performance I/O.

By locking the data into a binary format after the initial split, we guarantee that all subsequent modeling steps across different environments, R and Python, remain consistent.

Code 
# split dataset
set.seed(42)
df_split <- rsample::initial_split(df_all, prop = 1/2)
df_train <- rsample::training(df_split)
df_test  <- rsample::testing(df_split)

# write dataset as parquets
arrow::write_parquet(df_train, "../data/train.parquet")
arrow::write_parquet(df_test, "../data/test.parquet")

Data Preview

Below is a summary of the first 1000 rows of the processed training dataset. The Exposure column will be utilized as an offset in our subsequent modeling to account for varying policy durations.