statsmodels — Statistical Modeling and Econometrics in Python

Introduction

statsmodels complements scikit-learn by focusing on classical statistical inference rather than prediction. It provides detailed model summaries with coefficients, standard errors, p-values, and confidence intervals — the output statisticians and economists expect from tools like R or Stata.

What statsmodels Does

• Fits linear and generalized linear models with comprehensive diagnostic output
• Implements time-series analysis including ARIMA, VAR, state-space models, and seasonal decomposition
• Provides nonparametric methods like kernel density estimation and lowess smoothing
• Runs hypothesis tests (t-test, F-test, Granger causality, unit root tests)
• Generates publication-ready regression tables and diagnostic plots

Architecture Overview

statsmodels follows a model-fit-results pattern. You specify a model class (OLS, Logit, ARIMA), call .fit() to estimate parameters, and receive a results object with properties for coefficients, residuals, information criteria, and statistical tests. Under the hood, estimation uses scipy.optimize and numpy linear algebra routines.

Self-Hosting & Configuration

• Install via pip: pip install statsmodels
• Depends on NumPy, SciPy, pandas, and patsy for formula-based model specification
• Use R-style formulas: sm.OLS.from_formula("y ~ x1 + x2", data=df)
• Configure optimizer parameters and covariance estimators per model
• Works in Jupyter notebooks with rich HTML output for model summaries

Key Features

• Comprehensive model summaries matching R/Stata output with AIC, BIC, R-squared, and residual diagnostics
• Time-series toolbox with ARIMA, SARIMAX, VAR, and exponential smoothing
• Robust covariance estimators (HC0-HC3, HAC, clustered) for correct inference under heteroscedasticity
• Mixed-effects models for hierarchical and panel data
• Survival analysis with Kaplan-Meier and Cox proportional hazards

Comparison with Similar Tools

• scikit-learn — focused on prediction accuracy; statsmodels provides inference statistics (p-values, confidence intervals)
• R (stats package) — the gold standard for statistical computing; statsmodels brings similar functionality to the Python ecosystem
• SciPy (scipy.stats) — provides individual tests; statsmodels offers full model estimation and diagnostics
• linearmodels — extends statsmodels with panel data and IV models; statsmodels covers the broader foundation

FAQ

Q: When should I use statsmodels instead of scikit-learn? A: Use statsmodels when you need to understand relationships (coefficients, significance, confidence intervals) rather than just predict outcomes.

Q: Does statsmodels support regularized regression? A: Yes. OLS and GLM classes support elastic net regularization via fit_regularized(), though scikit-learn may be more convenient for pure prediction tasks.

Q: Can I use statsmodels for time-series forecasting? A: Yes. ARIMA, SARIMAX, and state-space models are well-implemented with automatic parameter selection helpers.

Q: How does the formula API work? A: Use patsy-style formulas like "y ~ x1 + x2 + x1:x2" to specify models declaratively from a DataFrame, similar to R.

Sources

• https://github.com/statsmodels/statsmodels
• https://www.statsmodels.org