Synthetic South African health insurance data with a known data-generating process, for learning and testing pricing models. Every parameter can be changed, and the true expected frequency and severity are returned, so you can check how close a fitted model gets to the truth.
The two Kaggle datasets below are built with this package
(data-raw/kaggle.R):
simulate_portfolio(1e5, seed = 123, claims = TRUE)simulate_panel(2e5, years = 5, seed = 2026)# install.packages("remotes")
remotes::install_github("Sbusiso-Mdingi/healthsim")library(healthsim)
# One year of 100,000 policies
pf <- simulate_portfolio(1e5, seed = 123)
# The same portfolio design with a heavier smoking loading and no frailty
pf2 <- simulate_portfolio(1e5, seed = 1,
params = portfolio_params(freq_smoker = 0.5, frailty_var = 0))
# Policies plus one row per claim
x <- simulate_portfolio(1e4, seed = 1, claims = TRUE)
head(x$claims)
# Five years of an open cohort starting with 50,000 members
pn <- simulate_panel(5e4, years = 5, seed = 1)
# Faster claims inflation and higher lapses
pp <- panel_params(claims_inflation = 0.10, lapse_base = 0.15)
pn2 <- simulate_panel(5e4, years = 5, seed = 1, params = pp)
# Compare a fitted model with the truth
fit <- glm(ClaimCount ~ Age + Smoker + Plan + offset(log(Exposure)),
family = poisson, data = set_levels(pf))
truth <- true_frequency(pf)portfolio_params() and panel_params() list
every parameter with its default; see ?portfolio_params and
?panel_params.
Principal members aged 18 to 85 in the nine provinces, on three benefit tiers (Essential, Standard, Comprehensive). Claim counts are Poisson given a gamma frailty (negative binomial overall), with log-linear effects of age, maternity, BMI, smoking, children, province and plan. Each claim’s cost is a lognormal-Pareto mixture scaled by a log-linear severity mean. Amounts are in rand.
An open cohort followed year by year. Members age, gain weight, quit smoking and have children; new, younger members join; members lapse at the end of a year (more in the first year and among non-claimants) or die during it. Frailty has a lasting part, so claims history predicts future claims. Daily steps and sleep efficiency, joined by a Clayton copula, lower claim frequency, but they are only recorded for wellness programme members, who are more active than the rest. Cost per claim inflates each year while annual limits stay fixed, so more members reach the limit over time.
All data is artificial. The parameters are illustrative assumptions, not estimates from any insurer’s data.
citation("healthsim")