Python notes

pip install scelo

Python ≥ 3.9, numpy + pandas only in the core; the extras add statsmodels/scipy, lifelib, matplotlib, parquet/Excel I/O. The package lives at packages/scelo-py and runs as-is on the CPython Scelo IDE bundles.

Idioms

The accessor. Importing scelo registers df.sc, so every frame-first function is also a method — handy for people who chain:

import scelo as sc
df.sc.profile()
df.sc.clean("all")
df.sc.triangle().sc.mack()

Tables are DataFrames. A Table subclasses pandas.DataFrame; slicing keeps the title, basis and notes, .df strips them, .note("reviewed") appends one. In Jupyter, a Table renders as the usual HTML frame with its notes beneath.

Results are dataclasses. mack(...) → a ReservingResult (.table .ibnr .se .detail), glm(...) → a GLMResult (.coef .predict() .relativities()), wmtr(...) → a WmtrResult (.table .survival .drivers), each printing a one-line headline before its table.

Charts return figures. Every plot_* returns the figure (no pyplot state): show it by being the last expression in a cell, save it with fig.savefig(...) or sc.save_figure(fig, "out.png").

Optional dependencies degrade, loudly or gracefully — never silently wrongly. Without statsmodels the GLM falls back to a numpy IRLS tested to agree to 1e-5 (the model header names the engine used). Without scipy the risk module uses numpy fallbacks. Without matplotlib, plot_* says charts need matplotlib: pip install matplotlib. Without lifelib, lifelib_run points you at pip install "scelo[life]" — and basicterm / scr_life / csm never needed it.

Determinism. Everything simulated is seeded by default — bootstrap, hull_white, aggregate_loss, and the WMTR engine (whose Mulberry32 stream is bit-exact with the IDE). Rerunning a script reproduces the pack.

Testing

pip install "scelo[dev]" && pytest

Ninety-plus tests: the IDE's golden actuarial-table identities, the hand-computed descriptive statistics, Mack's published RAA figures, the TypeScript-generated WMTR fixture, numpy-vs-statsmodels GLM parity for five families, and every cleaning rule. tests/make_golden.py regenerates the cross-language fixture the R suite checks against — run it whenever a function's semantics change, and commit both together.