crabbymetrics

Rust-backed econometrics 🦀🔢 models with a scikit-adjacent Python API. Focus: extremely low runtime dependency footprint, simple NumPy-facing estimators, robust standard errors, and bootstrap support where it fits the estimator.

Features

• Linear, IV, and panel causal estimators: OLS, Ridge, FixedEffectsOLS, TwoSLS, HorizontalPanelRidge, SyntheticControl, SyntheticDID, AugmentedBalancing, MatrixCompletion, InteractiveFixedEffects
• Common panel causal API: HorizontalPanelRidge, SyntheticDID, AugmentedBalancing, and MatrixCompletion use balanced outcome and absorbing treatment matrices, then expose ATT, counterfactuals, treatment effects, event-study summaries, and group means through summary()
• Regularized and likelihood estimators: ElasticNet, Logit, Multinomial Logit, Poisson
• Moment and semiparametric estimators: GMM, BalancingWeights, EPLM, AverageDerivative, PartiallyLinearDML, AIPW
• Shared robust covariance options for the main linear estimators: vanilla, HC1, Newey-West, and cluster
• Weighted fits for OLS, Ridge, FixedEffectsOLS, and TwoSLS
• ElasticNet spans the ridge and lasso corners: use l1_ratio=0.0 for ridge-style shrinkage and l1_ratio=1.0 for lasso-style shrinkage
• PCA and KernelBasis for feature engineering before regression-style estimation
• Optimizers namespace exposing LBFGS, BFGS, NonlinearConjugateGradient, Gauss-Newton least squares, and SimulatedAnnealing
• fit, predict, summary, and bootstrap where meaningful for the estimator

Install

This package is built with pyo3/maturin and ships as native wheels.

PyPI: https://pypi.org/project/crabbymetrics/

uv pip install crabbymetrics

Example

import numpy as np
from crabbymetrics import OLS

x = np.random.randn(200, 3)
beta = np.array([1.0, -2.0, 0.5])
y = 0.3 + x @ beta + np.random.randn(200) * 0.1

model = OLS()
model.fit(x, y)
print(model.summary())

Anytime-valid OLS inference is available from OLS.summary(...):

import numpy as np
import crabbymetrics as cm

rng = np.random.default_rng(1)
n = 100
x = rng.normal(size=(n, 3))
x = x - x.mean(axis=0)
trt = rng.choice([0.0, 1.0], size=n)
y = (
    1.0
    + 1.4 * x[:, 2]
    + 2.3 * trt
    + 2.0 * x[:, 0] * trt
    + 3.0 * x[:, 1] * trt
    + rng.normal(size=n)
)
design = np.column_stack([x, trt, x * trt[:, None]])

model = cm.OLS()
model.fit(design, y)
g_star = cm.optimal_g(n, design.shape[1] + 1, alpha=0.05)
summary = model.summary(vcov="vanilla", anytime_valid=True, g=g_star)
print(summary["p_value"])
print(summary["confint"])

Panel causal estimators take matrices directly rather than long data frames:

import numpy as np
import crabbymetrics as cm

Y = np.random.randn(20, 12)
W = np.zeros_like(Y)
W[15:, 8:] = 1.0  # absorbing treatment matrix

model = cm.SyntheticDID()
model.fit(Y, W)
out = model.summary()
print(out["att"], out["event_study"].keys(), out["group_means"].keys())

BalancingWeights remains available as the lower-level calibration/reweighting API, but the paved panel path is estimator-first.

AugmentedBalancing.fit(Y, W, outcome_model=None) composes supplied untreated-outcome predictions with unit or unit-and-time residual balancing. Constructor options select cohort versus individual unit-weight targets and raw versus residualized weight fitting.

The Augmented Balancing vignette shows a staggered-adoption workflow with a matrix-completion nuisance surface, estimator comparisons, event-time plots, and weight-target diagnostics. The class reference documents every constructor option and summary() field.

The direct optimizer wrappers live under Optimizers and follow a lightweight scipy-style interface:

import numpy as np
from crabbymetrics import Optimizers

def objective(theta):
    return float((theta[0] - 1.0) ** 2 + 2.0 * (theta[1] + 2.0) ** 2)

def gradient(theta):
    return np.array([2.0 * (theta[0] - 1.0), 4.0 * (theta[1] + 2.0)])

result = Optimizers.minimize_lbfgs(objective, np.array([4.0, 3.0]), gradient)
print(result["x"], result["fun"])

Benchmarks

The latest cross-library runtime snapshot is checked in as benchmarks/runtime_comparison.png.

This benchmark used synthetic problems with p=5, sample sizes from 10^3 to 10^6, fit-only timing, and a 45-second per-fit timeout.

• OLS is competitive already and was faster than both scikit-learn and statsmodels at n=10^6.
• Poisson beats statsmodels comfortably but still trails scikit-learn at larger n.
• Logit and especially MultinomialLogit are the main performance gaps to close before adding more iterative GLM-style estimators.

Development

Create and populate the project virtual environment, then build the extension into that venv.

uv sync
uv run maturin develop

uv run maturin develop is sufficient for rebuilding and reinstalling the package in .venv once the environment exists. If you change Python dependencies or the pyproject.toml metadata, run uv sync again first.

Package versioning is sourced from Cargo.toml; the Python package metadata is dynamic.

To release to PyPI, use the Build wheels GitHub Actions workflow manually from master and enter the next version without the leading v (for example 0.6.1). The workflow bumps Cargo.toml, commits to master, creates the vX.Y.Z tag, runs tests, builds wheels and an sdist, publishes the GitHub Release, and publishes to PyPI. The older commit_tag_release.sh script remains as a local fallback for manually tagging the current Cargo.toml version.

Rendered examples and API docs live under docs/. Rebuild the site with uv run quarto render docs. For docs work, install the docs extra first: uv sync --extra docs.

Wheels

Wheels are platform-specific and included in GitHub releases. See the releases tab.