analyze-model-data

Analyze Model Data

Produce a reproducible evidence trail from raw data to model-ready outputs.

Workflow

1. Preserve the raw files and identify schema, units, missingness, duplicates, and target leakage risks.
2. Run scripts/profile_csv.py for CSV inputs, then inspect domain-specific anomalies manually.
3. Define the analysis question and evaluation metric before choosing transformations or models. For entity-by-time planning data, call algorithms.data_diagnostics.panel_diagnostics (or an equivalent registered implementation) to quantify temporal dependence, cross-entity correlation, support/zero inflation, and a time-ordered holdout before choosing an uncertainty model.
4. Split train/test data before fitting imputers, scalers, encoders, or feature selectors.
5. Compare against a simple baseline and report uncertainty, not only point metrics.
6. Save cleaned data, analysis code, configuration, figures, and a machine-readable result summary.
7. Document every exclusion, imputation, transformation, random seed, and the evidence for or against independent sampling. Preserve a time-aware holdout or stress slice when the task contains future planning; never replace it with a random split merely because the random split scores better.

Guardrails

• Never overwrite raw data.
• Do not remove outliers solely because they weaken the result.
• Do not infer causality from association without an identification strategy.
• Use Python or R according to the project context; do not force one language.

Resources

Read references/analysis-standards.md before modeling. Use scripts/profile_csv.py for deterministic first-pass profiling.

Executable Contract

Use scripts/execute_skill.py with the shared runtime. Supply the raw_data, data_dictionary, and official_problem inputs, and produce every contracted output role: data_profile, data_quality_report, cleaning_actions, eda_findings, leakage_report, processed_data_manifest, and data_analysis. The profile and EDA findings must include the applicable panel/time diagnostics, not only row counts and missingness. Do not advance a project or register a data-audit artifact without its successful signed Skill run.