R Python differences

R vs Python differences

This page records where the R greybox package and the Python port agree numerically and where they intentionally differ. It complements Roadmap (which tracks what is ported) by documenting how closely the ported functions match.

How parity is checked

The Python repository ships a suite of cross-language comparison tests that run the R package (via rpy2) and the Python port on the same inputs and assert agreement to a numerical tolerance:

Test file	Covers
tests/test_r_python_compare.py	General ALM fitting, prediction, selection.
tests/test_alm_distributions_compare.py	ALM across the supported distributions.
tests/test_diagnostics_compare.py	outlier_dummy() vs outlierdummy().
tests/test_pointlik_compare.py	Point-wise likelihood.
tests/test_aid_compare.py	Demand identification (aid() / aid_cat()) vs R. See AID.

These require rpy2 and the R greybox package installed; they are skipped otherwise.

Verified to agree (to numerical tolerance)

Quantity	R	Python	Notes
Coefficients	coef(m)	m.coef	Across supported distributions.
Fitted values	fitted(m)	m.fitted
Scale / sigma	m$scale	m.scale	df_residual = n - k.
Outlier dummies	outlierdummy()	outlier_dummy()	Same ids and bounds.
Point likelihood	pointLik()	point_lik()
STI strengths	stick()$strength	stick().strength	Bit-identical to 6 d.p. on AirPassengers (seasonal 0.1061, trend 0.8613, irregular 0.0326). See EDA.
Smoothers	lowess(), supsmu()	lowess(), supsmu()	Match R to ~1e-15 (the Python smoothers are pybind11 ports of R's C/FORTRAN). See Smoothers.
Demand type	aid()$name, $type	aid().name, .type	Categorical decisions (type, stockout start/end, new/obsolete) match exactly. See AID.

Small last-digit differences can arise from optimiser tolerances (both use gradient-free optimisers but different implementations) and BLAS rounding; they are within the test tolerances. The aid() information-criterion values agree only to about 1e-3 (Python nlopt vs R nloptr), but the resulting categorical classifications are identical.

Intentional convention differences

These are deliberate and will not be "fixed" — they reflect either a language idiom or a correction made during the port.

Topic	R	Python	Why
MPE / MAPE scale	fraction	fraction	The Python port returns these as a fraction (not ×100), matching R. Earlier Python versions returned percentages; aligned as of the port's parity work. See measures.
asymmetry centring	centred at C = 0	centred at C = 0	Python's asymmetry in measures() is centred at zero (not at the mean of the errors), matching R.
Distribution count	26 distributions	26 distributions	Same families; names match (dnorm, dlnorm, …).

Intentional interface differences

Topic	R	Python
Model fitting	formula interface: alm(y ~ x, data, ...)	scikit-learn style: ALM(...).fit(X, y). Use formula() to build X/y.
Naming case	camelCase functions (alm, xregExpander)	snake_case functions / PascalCase classes (ALM, xreg_expander). Full map on Home.
Result access	R lists with $ (m$coefficients)	objects with attributes (m.coef)
stick seasonal lags	lags defaults to frequency(y)	lags is required (Python arrays carry no frequency); raises ValueError if omitted. See EDA.
Time series input	ts objects carry frequency/start	plain numpy arrays / pandas Series; periodicity is passed explicitly.

Not in Python yet

Functions that are R-only (so there is nothing to compare) are listed on Roadmap.

References

See Resources for the full reference list.