insurance-datasets

Synthetic UK insurance datasets with known data generating processes. Built for testing pricing models.

When you are developing a GLM, a gradient boosted tree model, or any other pricing algorithm, you need data where you know what the right answer is. Real policyholder data is messy, access-controlled, and the true coefficients are unknown. This package gives you clean, realistic synthetic data where the true parameters are published — so you can verify your implementation produces the right coefficients.

Two datasets are available:

• Motor: UK personal lines motor insurance. 18 columns covering driver age, NCD, ABI vehicle group, area band, and more. Frequency and severity from a known Poisson-Gamma DGP.
• Home: UK household insurance. 16 columns covering property value, flood zone, construction type, subsidence risk, and security level. Same Poisson-Gamma structure.

Installation

uv add insurance-datasets

Or with pip:

pip install insurance-datasets

Requires Python 3.10+. Dependencies: numpy, pandas.

Quick start

from insurance_datasets import load_motor, load_home

motor = load_motor(n_policies=50_000, seed=42)
home = load_home(n_policies=50_000, seed=42)

print(motor.head())
print(motor.dtypes)

Motor dataset

from insurance_datasets import load_motor, MOTOR_TRUE_FREQ_PARAMS, MOTOR_TRUE_SEV_PARAMS

df = load_motor(n_policies=50_000, seed=42)

# One row per policy
# policy_id, inception_date, expiry_date, accident_year,
# vehicle_age, vehicle_group, driver_age, driver_experience,
# ncd_years, ncd_protected, conviction_points, annual_mileage,
# area, occupation_class, policy_type,
# claim_count, incurred, exposure

print(f"Claim frequency: {df['claim_count'].sum() / df['exposure'].sum():.3f}")
print(f"Mean incurred (claims only): {df[df['claim_count']>0]['incurred'].mean():.0f}")

Home dataset

from insurance_datasets import load_home, HOME_TRUE_FREQ_PARAMS, HOME_TRUE_SEV_PARAMS

df = load_home(n_policies=50_000, seed=42)

# policy_id, inception_date, expiry_date, accident_year,
# region, property_value, contents_value, construction_type,
# flood_zone, is_subsidence_risk, security_level, bedrooms,
# property_age_band, claim_count, incurred, exposure

z1 = df[df["flood_zone"] == "Zone 1"]
z3 = df[df["flood_zone"] == "Zone 3"]
ratio = (z3["claim_count"].sum() / z3["exposure"].sum()) / (z1["claim_count"].sum() / z1["exposure"].sum())
print(f"Zone 3 vs Zone 1 frequency ratio: {ratio:.2f}x  (true DGP implies exp(0.85) = 2.34x)")

Data generating processes

Motor: true parameters

The frequency model is Poisson with a log-linear predictor:

log(lambda) = log(exposure) + intercept
            + vehicle_group_coef * vehicle_group_value
            + driver_age_young * I(driver_age < 25)
            + driver_age_old * I(driver_age >= 70)
            + ncd_years_coef * ncd_years_value
            + area_B * I(area == 'B')  ...  + area_F * I(area == 'F')
            + has_convictions * I(conviction_points > 0)

from insurance_datasets import MOTOR_TRUE_FREQ_PARAMS, MOTOR_TRUE_SEV_PARAMS

print(MOTOR_TRUE_FREQ_PARAMS)
# {'intercept': -3.2, 'vehicle_group': 0.025, 'driver_age_young': 0.55,
#  'driver_age_old': 0.3, 'ncd_years': -0.12, 'area_B': 0.1, 'area_C': 0.2,
#  'area_D': 0.35, 'area_E': 0.5, 'area_F': 0.65, 'has_convictions': 0.45}

print(MOTOR_TRUE_SEV_PARAMS)
# {'intercept': 7.8, 'vehicle_group': 0.018, 'driver_age_young': 0.25}

The age effect blends linearly between 25 and 30 — ages 25-29 have a partial young-driver load. The severity model uses a Gamma distribution with shape=2 (coefficient of variation ~0.71).

Home: true parameters

from insurance_datasets import HOME_TRUE_FREQ_PARAMS, HOME_TRUE_SEV_PARAMS

print(HOME_TRUE_FREQ_PARAMS)
# {'intercept': -2.8, 'property_value_log': 0.18,
#  'construction_non_standard': 0.4, 'construction_listed': 0.25,
#  'flood_zone_2': 0.3, 'flood_zone_3': 0.85, 'subsidence_risk': 0.55,
#  'security_standard': -0.1, 'security_enhanced': -0.25}

print(HOME_TRUE_SEV_PARAMS)
# {'intercept': 8.1, 'property_value_log': 0.35,
#  'flood_zone_3': 0.45, 'contents_value_log': 0.22}

The home severity model uses Gamma with shape=1.5 (CV ~0.82). Household claims have higher severity volatility than motor — large flood and escape-of-water events create a fatter tail.

GLM coefficient recovery example

The point of known DGP parameters is that you can verify your model implementation. Here is a worked example with statsmodels:

import numpy as np
import statsmodels.api as sm
from insurance_datasets import load_motor, MOTOR_TRUE_FREQ_PARAMS

df = load_motor(n_policies=50_000, seed=42)

# Create binary conviction flag (matches the DGP)
df["has_convictions"] = (df["conviction_points"] > 0).astype(int)

# Area dummies (A is the base level)
for band in ["B", "C", "D", "E", "F"]:
    df[f"area_{band}"] = (df["area"] == band).astype(int)

features = [
    "vehicle_group", "ncd_years", "has_convictions",
    "area_B", "area_C", "area_D", "area_E", "area_F",
]
X = sm.add_constant(df[features])

model = sm.GLM(
    df["claim_count"],
    X,
    family=sm.families.Poisson(),
    offset=np.log(df["exposure"].clip(lower=1e-6)),
)
result = model.fit(disp=False)

print("Parameter recovery:")
print(f"  vehicle_group: fitted={result.params['vehicle_group']:.4f}  true={MOTOR_TRUE_FREQ_PARAMS['vehicle_group']:.4f}")
print(f"  ncd_years:     fitted={result.params['ncd_years']:.4f}  true={MOTOR_TRUE_FREQ_PARAMS['ncd_years']:.4f}")
print(f"  convictions:   fitted={result.params['has_convictions']:.3f}  true={MOTOR_TRUE_FREQ_PARAMS['has_convictions']:.3f}")

At 50k policies, slope estimates should be within a few percent of the true values. The intercept will differ if you omit any factor from the true DGP — the model absorbs the omitted effects into the intercept.

CatBoost example

If you are calibrating a gradient boosted tree and want to verify it respects the known monotonic relationships:

import numpy as np
import pandas as pd
from catboost import CatBoostRegressor, Pool
from insurance_datasets import load_motor

df = load_motor(n_policies=50_000, seed=42)
df["has_convictions"] = (df["conviction_points"] > 0).astype(int)

features = ["vehicle_group", "driver_age", "ncd_years", "has_convictions", "annual_mileage"]
X = df[features]
y = df["claim_count"] / df["exposure"]

pool = Pool(X, label=y)
model = CatBoostRegressor(iterations=500, depth=6, learning_rate=0.05, verbose=0)
model.fit(pool)

# Feature importances should rank ncd_years and driver_age highly
fi = pd.Series(model.get_feature_importance(), index=features).sort_values(ascending=False)
print(fi)

Design choices

Why Poisson-Gamma and not something more exotic? GLMs are still the industry standard for personal lines pricing. The DGP matches what a correctly specified production model would use. If you want to test Tweedie models, use the raw incurred as the response — the data supports it.

Why no missing values? This is a testing dataset. Missing value imputation is a separate problem. Mixing the two makes it harder to isolate algorithm correctness.

Why 50,000 policies as the default? Below about 10,000 policies, coefficient estimates become noisy enough that a correct implementation can look wrong. At 50,000 the estimates are stable. For quick unit tests, 1,000-5,000 is sufficient.

Why is the home DGP simpler than motor? The motor DGP has more interacting factors because motor pricing in the UK has more rating variables and more data to calibrate them. The home DGP reflects a less mature pricing environment where the dominant factors (flood, construction, subsidence) are fewer and cleaner.

Running the tests

Tests require statsmodels for the GLM recovery check:

uv add --dev statsmodels
uv run pytest

Licence

MIT. See LICENSE.