pmdarima-rs 0.1.0

A drop-in replacement for pmdarima — the same auto_arima, the same ARIMA, the same arguments and the same selected orders, with the Kalman filter rewritten in Rust.

Install

pip install pmdarima-rs

Prebuilt wheels for Linux, macOS and Windows, Python 3.10 through 3.14 from a single abi3 wheel per platform. No compiler needed, no Rust toolchain needed.

Switch

Change one import.

- import pmdarima as pm
+ import pmdarima_rs as pm

That is the whole migration. If you cannot edit the code that imports pmdarima — someone else's library, a notebook you were handed, a vendored script — alias it instead.

Use it

Everything works the way it does in pmdarima, because it is the same API.

import pmdarima_rs as pm

y = pm.datasets.load_wineind()
model = pm.auto_arima(y, seasonal=True, m=12, trace=True)

forecast, ci = model.predict(n_periods=12, return_conf_int=True)
print(model.summary())

The summary() table is statsmodels' own layout, and every figure in it — coefficients, standard errors, information criteria, the Ljung-Box and Jarque-Bera block — is checked against statsmodels in the test suite.

See the full summary output

Speed

Every row is checked for agreement before it is timed, so a fast wrong answer cannot appear in the table. pmdarima's filter is already compiled — it is Cython — so this is not "Python versus native"; the wins are algorithmic, and docs/DESIGN.md says what they are.

These are one machine's numbers and they are noisy: repeat runs move the totals by 15–20% and individual rows by more. What holds across runs is the order of magnitude and the agreement. Every measurement, and how to reproduce it, is in docs/BENCHMARKS.md.

Accuracy

The claim is not "similar results".

The tests are differential, not golden-file: they run pmdarima and this package on the same input and compare. Where the two genuinely differ — three places where we are more accurate, and a handful where a hard threshold makes the choice a coin flip — it is measured and written down in docs/CORRECTNESS.md rather than smoothed over.

For code you cannot edit

import pmdarima_rs
pmdarima_rs.install()      # before the first `import pmdarima`

import pmdarima            # now resolves to pmdarima_rs

install() puts a finder on sys.meta_path, so every submodule resolves too, including ones nothing has imported yet: import pmdarima.arima.utils or from pmdarima.datasets.wineind import load_wineind gets the very same module object this package exposes. isinstance checks and pickles therefore still work across the alias.

One attribute is deliberately not passed through: pmdarima.__version__ reports the pmdarima API level this package implements (2.1.1), because the code you cannot edit is exactly the code likely to gate on it. This package's own version stays available as pmdarima.__pmdarima_rs_version__ and as pmdarima_rs.__version__.

What is included

All of it. A test reads pmdarima's own __all__ for each module and fails on any name this package does not provide.

The estimators are real scikit-learn estimators when scikit-learn is installed, and use a local fallback base when it is not — so clone, get_params, set_output and metadata routing all work, without making scikit-learn a dependency.

Dependencies

Three, against pmdarima's eight.

matplotlib is imported lazily inside the plotting helpers, so it only has to be there if you call them.

Limitations

• method accepts the same nine solver names statsmodels does and rejects anything else with the same ValueError, but only 'lbfgs' is implemented; the other eight warn and fall back to it. It is pmdarima's default and the only one its own auto_arima uses.
• Five SARIMAX options raise NotImplementedError rather than being accepted and quietly ignored: simple_differencing, measurement_error, time_varying_regression, mle_regression=False and use_exact_diffuse. Each of them changes the model, so honouring the argument by ignoring it would report a different model's numbers under your specification. Every other SARIMAX keyword is either implemented or genuinely makes no difference to the likelihood, and an unrecognised one is a TypeError, as it is in statsmodels.
• Order selection agrees with pmdarima on real data (10/10) but can differ on series whose fitted MA roots sit on auto_arima's 0.99 rejection threshold, where two optimisers agreeing to nine digits still land on opposite sides of a hard cutoff. It is measured in the benchmark rather than asserted away; details in docs/CORRECTNESS.md.
• The stationary initial covariance is solved by squaring, which is far cheaper than the k² × k² factorisation it replaces but is still the largest fixed cost per likelihood evaluation — roughly a sixth of a seasonal call at m = 12. An O(r²) recursion using the ARMA autocovariances would remove most of that; it is the clearest remaining headroom and is not implemented.

Development

git clone https://github.com/Fatin-Ishraq/pmdarima-rs
cd pmdarima-rs
pip install maturin
maturin develop --release

pip install "pmdarima>=2.1.1" statsmodels pytest   # the reference to test against
pytest tests/ -q
cargo test --lib
python bench/bench.py

The test suite skips its differential tests when pmdarima is not installed, so it still runs without the reference — it just checks less.

Documentation

• docs/DESIGN.md — where the time went and what replaced it, what is compiled and what is not
• docs/CORRECTNESS.md — what is verified, how, and every known difference
• docs/BENCHMARKS.md — every measurement, and how to reproduce it
• pmdarima's own documentation applies unchanged

Licence

MIT. pmdarima is MIT (Taylor G. Smith and contributors); statsmodels is BSD-3. Portions of this package are ports of both, as noted in the module docstrings.