Recipe

The general regression recipe

Use when: you have a labeled dataset and need to predict a continuous numeric value

Goal: a tuned, honestly evaluated regression model with a documented baseline comparison

Taught in Which algorithm should I use?

1. 1

   Confirm exactly what is being predicted. Note the time window, if one matters. Note which features are only known after the fact.

   Why: the same leakage risk from classification applies here too, just for a number instead of a group.

2. 2

   Compute the error from always guessing the training set's mean target value. Do this before fitting anything more complex.

   Why: an RMSE or MAE number only makes sense next to this reference point.

3. 3

   Fit a regularized linear regression, a random forest, and a gradient-boosted model for regression. Run all three through the identical Pipeline and CV setup.

   Why: as with classification, different model families suit different relationships between features and target.

4. 4

   Pick the error metric that matches how errors actually matter here. RMSE punishes big errors harder than MAE does.

   Why: two models with similar R-squared can still have very different error patterns, depending on which metric is used.

5. 5

   Apply random or Bayesian search. Use it only on the single strongest candidate from the last step.

   Why: same reasoning as the classification recipe -- tuning time is best spent on the already-strongest starting point.

6. 6

   Check feature importance as a final sanity check. Then save the entire fitted Pipeline together.

   Why: the same mismatch risk applies, whether the target is a group or a number.