name: statsmodels-python
description: >-
Use for writing, reviewing, debugging, or interpreting Python statistical models with statsmodels, including formulas, regression, GLM, time series, robust covariance, diagnostics, prediction intervals, and inference. Do not use for sklearn prediction pipelines or Bayesian posterior workflows.
argument-hint: "[statsmodels Python task, code, contract, or failure]"
---
# statsmodels Python
Use the smallest explicit execution contract that preserves semantics. Inspect
the installed version before relying on a drifting signature. State ownership,
input and output shapes, ordering, failure behavior, and the verification command
before writing substantial code.
## Object and execution model
| Object | Meaning | Boundary |
|---|---|---|
| `model specification` | A likelihood or estimating-equation family plus design. | Formula coding, intercept, link, and covariance assumptions define estimands. |
| fitted result | Parameters, covariance, diagnostics, and prediction methods. | Check convergence and data rows retained before inference. |
| design matrix | Encoded predictors with column semantics. | Categorical levels and transformations must match prediction data. |
| covariance estimator | An uncertainty model such as classical, HC3, cluster, or HAC. | Choose it from dependence and sampling, not coefficient preference. |
| prediction frame | Mean and possibly observation uncertainty at new exog. | Mean confidence and observation intervals answer different questions. |
Read [the object model](references/object-model.md) when the task mixes two
objects or crosses an execution boundary.
## Workflow
1. Define estimand, outcome family, sampling unit, clustering/order, and missing-data policy.
Build and inspect the formula/design matrix and retained observations.
Fit with the intended likelihood and covariance estimator.
Check convergence, residual, influence, and specification diagnostics.
Generate predictions with explicit exog and the correct uncertainty target.
## Decision rules
- State estimand, response distribution, link, design formula, intercept, weights, and dependence assumptions before fitting.
Choose classical, heteroskedasticity-robust, cluster-robust, or HAC covariance from the sampling process.
Inspect dropped rows, rank, convergence, residuals, influence, and model-specific diagnostics before interpreting p-values.
Preserve categorical levels and formula transformations when constructing prediction data.
Distinguish confidence intervals for the conditional mean from prediction intervals for new observations.
Treat association and conditional estimates as causal only when the identification design independently supports that claim.
If a required fact is unknown, inspect the target code, installed signature,
schema, shape, or lifecycle owner. Do not replace an unknown with a permissive
fallback. See [the decision guide](references/decision-guide.md).
## Complex solution routes
Load only the matching section of
[the evaluated recipes](references/recipes-solutions.md):
- `statsmodels.formula-ols-robust` and `statsmodels.verify-design-and-covariance`: fit a categorical regression with robust covariance.
statsmodels.prediction-summary-frame and statsmodels.verify-interval-semantics: return distinct prediction uncertainty targets.
statsmodels.hac-ordered-fit and statsmodels.verify-time-and-lag-contract: use HAC covariance only on verified time ordering.
Recipes are anchors, not blind templates. Preserve their named invariants and
adapt types and names only after inspecting the actual boundary.
## Verification contract
- Test observable behavior, not the presence of API tokens.
- Exercise empty, singleton, malformed, and failure inputs when the operation
accepts them.
- Assert shape, dtype or type, ordering, ownership, and error semantics where
they are part of the contract.
- Keep external I/O deterministic with injected clocks, transports, processes,
files, random state, or test doubles.
- Run the narrow test first, then the relevant project suite. Do not declare
completion when warnings, background failures, convergence flags, or cleanup
errors remain unexplained.
Use [the verification matrix](references/verification.md) for completion checks.
## Failure routing and adaptation
Classify a failure before changing code: input-contract failures require a
precise rejection; environment or version failures require inspection; execution
failures require lifecycle, convergence, or cleanup evidence; invariant failures
require a semantic correction. Do not relax a check, coerce a value, broaden a
failure handler, or materialize data merely to make the symptom disappear.
When adapting a recipe:
1. Match its objects, ownership, execution timing, and output contract to the task.
2. Preserve every branch condition and completion check while changing domain names.
3. Add the project's real empty, malformed, duplicate, cancellation, precision, or
boundary case before removing any guard.
4. If the installed API differs, inspect the signature and primary documentation,
then update implementation, test, and authoring evidence together.
## Version grounding
Inspect the installed package and signature when editing an existing project.
Treat examples here as verified anchors for the version recorded by the Foundry,
not as permission to overwrite a repository's compatibility policy. When current
behavior differs, preserve the project target and update tests and authoring
evidence together.
## Completion
Complete the task only when the implementation preserves the declared object
model, no accidental materialization or lifetime extension was introduced, all
failures are surfaced at the correct boundary, and deterministic tests prove the
critical behavior. Report any environment or version fact that could not be
verified.
1---2 name: statsmodels-python3 description: >-4 Use for writing, reviewing, debugging, or interpreting Python statistical models with statsmodels, including formulas, regression, GLM, time series, robust covariance, diagnostics, prediction intervals, and inference. Do not use for sklearn prediction pipelines or Bayesian posterior workflows.5 argument-hint: "[statsmodels Python task, code, contract, or failure]"6 ---78 # statsmodels Python910 Use the smallest explicit execution contract that preserves semantics. Inspect11 the installed version before relying on a drifting signature. State ownership,12 input and output shapes, ordering, failure behavior, and the verification command13 before writing substantial code.1415 ## Object and execution model1617 | Object | Meaning | Boundary |18 |---|---|---|19 | `model specification` | A likelihood or estimating-equation family plus design. | Formula coding, intercept, link, and covariance assumptions define estimands. |20| `fitted result` | Parameters, covariance, diagnostics, and prediction methods. | Check convergence and data rows retained before inference. |21| `design matrix` | Encoded predictors with column semantics. | Categorical levels and transformations must match prediction data. |22| `covariance estimator` | An uncertainty model such as classical, HC3, cluster, or HAC. | Choose it from dependence and sampling, not coefficient preference. |23| `prediction frame` | Mean and possibly observation uncertainty at new exog. | Mean confidence and observation intervals answer different questions. |2425 Read [the object model](references/object-model.md) when the task mixes two26 objects or crosses an execution boundary.2728 ## Workflow2930 1. Define estimand, outcome family, sampling unit, clustering/order, and missing-data policy.312. Build and inspect the formula/design matrix and retained observations.323. Fit with the intended likelihood and covariance estimator.334. Check convergence, residual, influence, and specification diagnostics.345. Generate predictions with explicit exog and the correct uncertainty target.3536 ## Decision rules3738 - State estimand, response distribution, link, design formula, intercept, weights, and dependence assumptions before fitting.39- Choose classical, heteroskedasticity-robust, cluster-robust, or HAC covariance from the sampling process.40- Inspect dropped rows, rank, convergence, residuals, influence, and model-specific diagnostics before interpreting p-values.41- Preserve categorical levels and formula transformations when constructing prediction data.42- Distinguish confidence intervals for the conditional mean from prediction intervals for new observations.43- Treat association and conditional estimates as causal only when the identification design independently supports that claim.4445 If a required fact is unknown, inspect the target code, installed signature,46 schema, shape, or lifecycle owner. Do not replace an unknown with a permissive47 fallback. See [the decision guide](references/decision-guide.md).4849 ## Complex solution routes5051 Load only the matching section of52 [the evaluated recipes](references/recipes-solutions.md):5354 - `statsmodels.formula-ols-robust` and `statsmodels.verify-design-and-covariance`: fit a categorical regression with robust covariance.55- `statsmodels.prediction-summary-frame` and `statsmodels.verify-interval-semantics`: return distinct prediction uncertainty targets.56- `statsmodels.hac-ordered-fit` and `statsmodels.verify-time-and-lag-contract`: use HAC covariance only on verified time ordering.5758 Recipes are anchors, not blind templates. Preserve their named invariants and59 adapt types and names only after inspecting the actual boundary.6061 ## Verification contract6263 - Test observable behavior, not the presence of API tokens.64 - Exercise empty, singleton, malformed, and failure inputs when the operation65 accepts them.66 - Assert shape, dtype or type, ordering, ownership, and error semantics where67 they are part of the contract.68 - Keep external I/O deterministic with injected clocks, transports, processes,69 files, random state, or test doubles.70 - Run the narrow test first, then the relevant project suite. Do not declare71 completion when warnings, background failures, convergence flags, or cleanup72 errors remain unexplained.7374 Use [the verification matrix](references/verification.md) for completion checks.7576 ## Failure routing and adaptation7778 Classify a failure before changing code: input-contract failures require a79 precise rejection; environment or version failures require inspection; execution80 failures require lifecycle, convergence, or cleanup evidence; invariant failures81 require a semantic correction. Do not relax a check, coerce a value, broaden a82 failure handler, or materialize data merely to make the symptom disappear.8384 When adapting a recipe:8586 1. Match its objects, ownership, execution timing, and output contract to the task.87 2. Preserve every branch condition and completion check while changing domain names.88 3. Add the project's real empty, malformed, duplicate, cancellation, precision, or89 boundary case before removing any guard.90 4. If the installed API differs, inspect the signature and primary documentation,91 then update implementation, test, and authoring evidence together.9293 ## Version grounding9495 Inspect the installed package and signature when editing an existing project.96 Treat examples here as verified anchors for the version recorded by the Foundry,97 not as permission to overwrite a repository's compatibility policy. When current98 behavior differs, preserve the project target and update tests and authoring99 evidence together.100101 ## Completion102103 Complete the task only when the implementation preserves the declared object104 model, no accidental materialization or lifetime extension was introduced, all105 failures are surfaced at the correct boundary, and deterministic tests prove the106 critical behavior. Report any environment or version fact that could not be107 verified.