linear-regressionlisted

# Linear Regression ## When to use this skill Use for **continuous outcome prediction or explanation** where interpretability matters. Triggers: - "Predict revenue / session time / NPS" - "What drives <continuous metric>?" - "Linear regression for…" - "OLS" - "Explain the variation in…" For binary outcomes use `logistic-regression`. For time-to-event use `survival-analysis`. For pure prediction with non-linear effects, fit linear first as baseline, then suggest gradient boosting. ## Required inputs | Input | Why it matters | |---|---| | Continuous target | What you're predicting (numeric, ideally not heavily skewed) | | Feature set | Predictors | | Observation grain | Per user / per session / per geography | | Train/test strategy | Temporal split for production, random for exploratory | | Goal | Pure prediction, coefficient interpretation, or both? | ## Workflow 1. **Audit the data** (`data-quality-audit` skill). 2. **Inspect the target distribution.** ```python import matplotlib.pyplot as plt df[target].hist(bins=50) df[target].describe() ``` - Symmetric, finite variance → OK for OLS - Heavily right-skewed (revenue, time-on-page) → consider `log(1 + target)` transformation - Heavy-tailed with extreme outliers → robust regression or winsorize 3. **Check for outliers** at p99 and p99.9. Decide: - Cap at p99 (winsorize) - Drop with documented justification - Keep and use robust regression 4. **Verify no leakage** (same checks as logis