R notes

install.packages("scelo", repos = c("https://intelligentactuaries.com/r", getOption("repos")))
library(scelo)

R ≥ 4.1, and the package imports nothing beyond stats, utils and tools — it installs on a bare R, including the one Scelo IDE bundles. (CRAN requires an open-source licence; this package is distributed from the repository.) jsonlite + curl unlock the swarm client, statmod Tweedie GLMs, reticulate the lifelib bridge.

Idioms

Everything is sc_. Type the prefix and let completion show the map — the R analogue of sc. in Python. Pipes read naturally:

df |> sc_clean() |> sc_triangle() |> sc_mack()

A scelo_table is a data.frame. Subsetting keeps the title, basis and notes; sc_df(x) strips them; sc_notes(x), sc_basis(x), sc_title(x) read them; sc_note(x, "reviewed") appends. Printing truncates at 60 rows — options(scelo.print_rows = 200) to raise it.

Results are lists with S3 print methods. sc_mack(tri) prints its headline then the table; the pieces are $table, $ibnr, $se, $detail. summary(m) gives the one-row frame; predict(m, new) works on a scelo_glm.

Back-tick the actuarial glyphs. sc_factors() names its columns the way the quantities are written — t$`äx`, t$`A¹x:10`.

Seeds behave three ways, worth knowing in a long session: sc_bootstrap() and sc_hull_white() seed themselves and restore your session RNG; sc_aggregate_loss(method = "mc"), sc_simulate_losses() and sc_reservoir() call set.seed() and leave it; the WMTR engine uses its own Mulberry32 stream and never touches R's RNG at all.

Base R only, deliberately. Sorting uses radix order where determinism matters (locale-independent, so results match Python byte-for-byte); dates parse to UTC; text I/O is UTF-8 throughout.

Testing

install.packages("testthat")
# from packages/scelo-r:
testthat::test_dir("tests/testthat")

The suite reads tests/testthat/fixtures/py_golden.json — values computed by the Python package — and checks life tables, commutation functions, factors, premiums, the RAA Mack table, discount curves, Smith–Wilson, Panjer, credibility, GLM coefficients, the cleaning plan for the dirty sample and more to 1e-9; the WMTR engine is checked against the same TypeScript fixture the Python package uses, and dev/build.sh runs the full R CMD check pipeline.