Capstone: a regression project

Same data, better target. late was always minutes > 30 — somebody's threshold — so predicting the number keeps what the threshold discarded, and you can apply any cut you like afterwards.

It also lets you report in minutes, which is a unit the stall owner already thinks in. That matters more than it sounds.

1. Split — and note what changes

X_train, X_test, y_train, y_test = train_test_split(
    X, deliveries["minutes"], test_size=0.25, random_state=0)

No stratify: the target is continuous, so there are no classes to balance. (If you needed it — a very skewed target, say — you'd bin the target and stratify on the bins.)

2. Baseline

DummyRegressor(strategy="mean") scores R² of essentially 0 by construction, which is R²'s quiet advantage: the baseline is built into the scale, so you don't have to fit one to read the number. Do it anyway for the MAE, which is 4.91 minutes — that's what "no model" costs you.

3. Pipeline and search

pipe = Pipeline([("pre", pre), ("m", Ridge())])

GridSearchCV(pipe, {"m__alpha": np.logspace(-2, 3, 12)},
             cv=5, scoring="r2").fit(X_train, y_train)

best alpha    5.337
cv R²         0.5871

Ridge rather than plain LinearRegression because the penalty costs nothing when it isn't needed and protects you when it is (lesson 15). Log scale on alpha, because its effect is multiplicative (lesson 23).

Module 4 already established the important negative results here: polynomial features make it worse, and a 200-tree forest scores 0.5198 against linear's 0.5785. This process really is a weighted sum, and it's worth having checked.

4. Diagnose before you report

resid = y_test - model.predict(X_test)

residual mean    −0.0236        centred
worst miss        9.93 minutes
within 5 min      80.0%

Plot residuals against predictions and you get a flat, shapeless band — no curve, no fan, no drift (lesson 17). The straight-line assumption holds, which independently corroborates why the forest couldn't beat it.

Do this before you report the score. A good R² with a curved residual plot means you have a model that works on average and fails systematically somewhere specific, and the average will not tell you where.

5. Report in units people use

TEST R²      0.5967
TEST MAE     3.087 minutes
within 5 min 80.0%

Three numbers doing three different jobs:

• R² — for a technical reader, with its baseline built in.
• MAE — the typical miss, in minutes.
• "80% within five minutes" — the one for the stall owner. It's a promise they can check against reality tomorrow, and it converts a model score into an operational commitment.

That third form is worth constructing deliberately for every regression project. Pick a tolerance that matters to the business and report the fraction inside it.

6. And connect it back

You can recover the classification decision from the regression at any threshold: predicted_minutes > 30. That's often the better architecture — one model, one number, and the operational cut kept as a separate, adjustable decision rather than baked in at training time.

It also means changing the promise from 30 minutes to 25 requires no retraining at all.

The checklist

1. Predict the number if you have it; you can always threshold later.
2. Split (no stratify), baseline, pipeline.
3. Tune alpha on a log scale by cross-validation.
4. Plot the residuals before you believe the score.
5. Report R², an error in real units, and a tolerance fraction.