FabianCornejo/econometric-skills

Agent skills for Claude focused on applied econometrics.
## econometrics
An econometric agent for Claude that works as a consultant, a teacher or a reviewer. It turns a plain question ("do interest rates hurt my sales?", "for which customers does the discount work?", "what beta should go in this WACC?") into an estimated, diagnosed and defensible answer, explains the method behind it, or audits someone else's analysis. Everything runs in Python (statsmodels, linearmodels, arch, scikit-learn) through tested scripts, and Claude answers in the user's language.
### Three modes, one tested core
| Mode | When | What you get |
| Consultant (default) | Estimate, test, forecast, evaluate a policy | A study card (estimand, design, identifying assumption), the analysis with diagnostics and robustness, a report and a reproducible script |
| Teacher | "Explain", "teach", "why does...", a class, an exam | A method card: intuition, model, assumptions, estimator, derivation, a seeded simulation that shows the assumption failing, code, diagnostics, interpretation, pitfalls, and an exercise with its solution |
| Reviewer | Review, audit or referee a paper, report or expert opinion | An opinion with the claim, a verdict, findings by severity, and the skill function that runs each check |
The mode changes the form of the answer, never the computations: every number comes from the same scripts.
### What it covers
| Area | Methods | Script |
| Statistics foundations | Describing data (skewness, kurtosis in both conventions, outliers), one- and two-sample t (Welch by default, paired), variances, proportions (Wilson, exact, Newcombe) and sample size, goodness of fit and normality, contingency tables with Fisher, Cramer's V and a Simpson's-paradox check, rank tests with Hodges-Lehmann intervals, correlations, power curves, multiple-testing adjustments | inference.py |
| Regression and diagnostics | OLS with full test batteries, robust/HAC SEs, influence, linear restriction tests (SSR form, restricted estimates), Prais-Winsten/Cochrane-Orcutt, J and encompassing tests for non-nested models, FGLS, prediction intervals, omitted-variable bias, Oster and Cinelli-Hazlett sensitivity | diagnostics.py, specification.py |
| Classical linear models | ANOVA and multiple comparisons, ANCOVA, robust and quantile regression, permutation tests, Box-Cox, lasso/ridge/PCR/PLS, post-double-selection | linearmodels.py |
| Estimation engine | Maximum likelihood from any likelihood, LR/Wald/LM, NLS, delta method, bootstrap, wild cluster bootstrap | estimation.py |
| Causal graphs | Minimal and optimal adjustment sets, good vs bad controls (confounders, mediators, colliders, instruments), tests of the graph's implications | dag.py |
| Causal inference | Matching, IPW, AIPW, double ML with omitted-confounder bounds and benchmarks, regression discontinuity, staggered DiD (Callaway-Sant'Anna) with timing checks and covariates (doubly robust, conditional parallel trends), the common causal covariates test, a TWFE hidden-linearity diagnostic, the Goodman-Bacon decomposition, cluster bootstrap at the assignment level, synthetic control, randomization inference for few treated units | causal.py |