How to use the Statsmodels Statistical Modeling agent skill

What the statsmodels skill makes the agent do

Without a skill, an agent writes statistical code from general knowledge. With this skill loaded, the agent follows guidance written for the statsmodels library: it fits classic models such as linear regression, logistic and Poisson regression, mixed and discrete choice models, and ARIMA or SARIMAX forecasts. It reports coefficients, confidence intervals, marginal effects and model comparisons.

The skill also asks the agent to check assumptions, including heteroskedasticity, autocorrelation and influential points. It states that a model fitting successfully does not prove causation or that the model is adequate, so the agent should flag design and assumption problems rather than present a clean fit as a finding.

• Regression: OLS, WLS, GLS, quantile regression
• Generalized linear models: logistic, Poisson, Gamma and similar families
• Discrete outcomes: binary, multinomial, count and ordinal models
• Time series: ARIMA, SARIMAX, stationarity tests and forecasts
• Model comparison with AIC/BIC and likelihood ratio tests

Install the statsmodels skill from the ZIP

Download the ZIP for this skill from our page and unzip it. The skill is a folder that contains SKILL.md and a references folder with several topic files. Place that folder in one of these locations, depending on who should use it: ~/.claude/skills/ for your personal use across projects, or .claude/skills/ inside a specific project folder to share it with that project.

If you use the Claude apps instead, you can upload the skill under Settings → Capabilities → Skills.

The Python libraries are separate from the skill. The source gives this command for a dedicated environment, with the pinned versions it was tested against. Run it only if you want the agent to execute the code, and it needs network access.

Check your Python and library setup

The skill requires Python 3.10 or newer. The tested stack needs Python 3.12 or newer, and the source reports native checks on Python 3.13. Plotting needs matplotlib, and predictive metrics need scikit-learn, but both are optional for basic fitting.

No API keys or credentials are needed. The network is used only for installation or documentation lookups.

• Python 3.10+ is the minimum; 3.12+ is what the tested stack needs
• matplotlib is optional, for plots
• scikit-learn is optional, for predictive metrics
• No API keys or credentials

Trigger the skill with natural requests

You do not need special syntax. Describe the analysis you want in plain language and mention your data file and the outcome you care about. The skill is meant to be used when you fit regression models, run generalized linear models, analyze discrete outcomes, forecast a time series, test assumptions, detect outliers, compare models, or produce publication-ready statistical tables.

Be specific about the design. The agent should use chronological splits for time series and grouped splits for repeated units, so tell it if your data has either structure.

• Fit a logistic regression of churn on tenure and plan type in data.csv, then check for influential observations and report odds ratios.
• Fit an OLS model of price on size and location, check heteroskedasticity, and refit with robust standard errors if needed.
• Forecast monthly sales with a SARIMAX model, check residuals with Ljung-Box, and evaluate on the last 12 months.
• Compare a Poisson model with a negative binomial model for visit counts and tell me which fits better using AIC.

Example workflow: a linear regression with diagnostics

The skill sets out a standard order for a linear regression. The agent explores the data, fits an initial OLS model, checks residuals, tests heteroskedasticity and autocorrelation, checks multicollinearity with VIF, identifies influential observations, and refits with robust standard errors if needed. It then interprets the coefficients and validates the model on a holdout set or with cross-validation.

You can ask the agent to follow this order and to explain each step in plain terms. Keep the reasoning and the decisions it made, such as any excluded rows, so you can review them.

• Explore data with plots and descriptive statistics
• Fit the initial OLS model and check residuals
• Test heteroskedasticity and autocorrelation
• Check VIF and influential observations
• Refit with robust SEs if needed, then validate

Safety notes: scripts, network and what to review

The skill runs scripts and uses the network, and its tool list includes Read, Write, Edit and Bash. Before you let the agent run code, read what it plans to execute and confirm that it is working on the data you intended. Installation needs network access, so run it in an environment you control.

The skill itself warns about limits. A successful fit is not proof of causation, and an agent may produce a plausible but wrong specification. For any decision that matters, have a human statistician review the code and assumptions.

• Review commands before the agent runs them, especially Bash calls
• Keep raw data and the installed environment separate from other projects
• Treat results as drafts until a qualified person has checked the design and assumptions

Tips for better results

Give the agent the full context: the outcome, the unit of observation, whether observations repeat over time or groups, and what question the model should answer. The skill says that the design and the assumptions matter more than the fit itself.

Ask for the diagnostics, not just the coefficients. Request confidence intervals, marginal effects for logistic models, and the reasons it chose a model. For time series, ask it to justify differencing and deterministic terms using plots and tests rather than a single p-value.

• State the research question and the study design up front
• Ask for confidence intervals alongside every estimate
• For causal questions, describe the study design first
• Ask the agent to log which rows were dropped and why