logistic-regressionlisted

# Logistic Regression ## When to use this skill Use for **binary outcome prediction** where interpretability matters as much as accuracy. Triggers: - "Model churn / conversion / fraud / adoption" - "Predict probability of X" - "Build a propensity model" - "What predicts <binary outcome>?" - "Logistic regression for…" For complex non-linear interactions, suggest a tree-based model after fitting logistic as the baseline. For continuous outcomes, use `linear-regression`. For time-to-event, use `survival-analysis`. ## Required inputs | Input | Why it matters | |---|---| | Binary target | What you're predicting (must be 0/1 or boolean) | | Feature set | Predictors, with clear definitions | | Observation grain | Per user / per session / per opportunity | | Time cutoffs | Feature window must end BEFORE outcome window starts (leakage check) | | Train/test strategy | Temporal split for production models, random split for exploratory | ## Workflow 1. **Audit the data first** (use `data-quality-audit` skill). 2. **Check class balance.** ```python print(y.value_counts(normalize=True)) ``` - Balanced (40-60%): standard logistic - Mild imbalance (10-40%): use `class_weight='balanced'` and look at PR-AUC, not ROC-AUC alone - Severe (< 5%): consider downsampling, but always evaluate calibration on the original distribution 3. **Verify no leakage.** The most common DS bug. Ask: - Is any feature derived from data *after* the prediction time? - Is the target ev