crabbymetrics

crabbymetrics logo

crabbymetrics is a Rust-backed econometrics library with a compact Python API. The docs are organized around the public API, a single binding crash course, an active First Course Ding translation track, supervised and semiparametric example pages, unsupervised transform notes, focused numerical ablations, and a small set of supporting internals pages.

Start Here

• API reference: verified public surface, summary schemas, optimizer catalog, and runtime smoke checks.
• First Course Ding: chapter-by-chapter translation plan and current implementation status for the Peng Ding notebooks.
• Binding Crash Course: OLS: the shortest end-to-end walkthrough of the Rust-to-Python wrapper pattern in this codebase.

First Course Ding

• Overview And TOC: planning page and chapter map for the full Ding translation pass.
• Translated so far: Chapter 1, Chapter 2, Chapter 3, Chapter 4, Chapter 9, Chapter 11, Chapter 12, Chapter 13, Chapter 21, and Chapter 23.
• The next obvious tranche is Chapters 15 through 20, where matching, sensitivity analysis, and RD helpers start to matter.

Supervised Learning Examples

• OLS: baseline linear regression with switchable vanilla and HC1 covariance estimators.
• Ridge: closed-form L2-regularized least squares with a scalar penalty or a penalty grid plus cross-validation.
• Fixed Effects OLS: partial out one-way or multi-way categorical fixed effects with within, then estimate slopes without an intercept.
• ElasticNet: regularized linear regression.
• Synthetic Control: simplex-constrained donor weighting for treated-versus-donor panel matching under latent-factor drift.
• Logit: binary logistic regression.
• Multinomial Logit: multiclass classification.
• Poisson: count regression, with model-based or QMLE sandwich inference available from summary(vcov=...).
• TwoSLS: instrumental variables regression, including multiple endogenous regressors and multiple excluded instruments.
• GMM: fit just-identified score equations, overidentified two-step IV moments, and stacked nuisance-parameter moments with the first-class GMM estimator.
• FTRL: online-style binary classification.
• MEstimator Poisson: callback-driven estimation matched against the built-in Poisson estimator.

Semiparametric Examples

• Balancing Weights: entropy and quadratic calibration weights for ATT-style reweighting, covariate shift, and domain adaptation.
• EPLM: the Robins-Newey partially linear E-estimator, implemented as a stacked-moment estimator for a scalar continuous treatment.
• Average Derivative: Oaxaca-Blinder, generalized IPW, and doubly robust average-derivative estimators in the Graham-Pinto style under a continuous-treatment working model.
• Double ML And AIPW: cross-fit ridge nuisance estimation for partially linear DML and binary-treatment AIPW.
• Richer Regression With Transformer Pipelines: a longer semiparametric-style example built on KernelBasis and PCA.

Unsupervised Learning Examples

• PCA And Kernel Basis: transformer examples for richer design matrices and nonlinear geometry.

Ablations

• Variance Estimators: cached Monte Carlo coverage experiments for OLS, Poisson, and GMM variance estimators under heteroskedasticity or overdispersion.
• Semiparametric Estimator Comparisons: cached Monte Carlo comparisons for EPLM versus partially linear DML under continuous-treatment misspecification, and for vanilla regression versus balancing weights versus AIPW under binary treatment.
• Bridging Finite And Superpopulation: cached Monte Carlo coverage comparisons for HC2, Ding’s closed-form correction, a raw-row bootstrap, and stacked GMM when the same adjusted estimator is evaluated against SATE and PATE.

Optimization

• Optimizers: direct optimizer usage for smooth likelihoods, rougher objective surfaces, and solver behavior comparisons.
• GMM With Optimizers: the lower-level notebook that motivated the first-class GMM estimator and still shows the residual-collection view directly.

Supporting Pages

• Binding Internals: Poisson: a deeper built-in-estimator walkthrough after the OLS crash course.
• Binding Internals: MEstimator: the callback-heavy bridge where Rust owns optimization and Python supplies the objective and scores.

Runtime Snapshot

Repo-level runtime comparison for OLS, Logit, Poisson, and MultinomialLogit across crabbymetrics, scikit-learn, and statsmodels. Lower is faster.

The benchmark figure is generated from the repo-level benchmark assets in benchmarks/ and gives a quick scale check for the estimators that overlap cleanly with mainstream Python baselines.

Notes

• api.qmd remains the main documentation page and renders to api.html.
• The site is a Quarto website, so shared navigation and search are generated under docs/.
• All pages are rendered with embedded resources so the checked-in HTML files remain self-contained.