Pipelines

• Estimator, transformer and pipeline abstractions
• ColumnTransformer for mixed numeric and categorical data
• Fitting on training data only: the discipline that prevents leakage

Models

• Linear and regularised models, and reading their coefficients honestly
• Trees, random forests and gradient boosting
• Class imbalance handling and threshold selection

Validation

• Cross-validation strategies including grouped and time-series splits
• Metric selection matched to the decision being made
• Grid and randomised search without overfitting the validation set

Delivery

• Persisting pipelines and versioning models
• Interpretability with permutation importance and SHAP
• Packaging an analysis so a colleague can rerun it unchanged