The supervised learning workflow, end to end

Most ML projects fail in framing and data, not modeling. The algorithm is the easy part. This is the loop that keeps you honest.

The loop

1. Frame the problem. What decision does this output drive? Classification or regression? What does a wrong answer cost (this picks your metric)?
2. Get and split the data before you touch it — train/validation/test, by time or group if needed (see leakage).
3. Build a baseline. A constant, a heuristic, or a logistic regression. This is the bar every fancier model must beat.
4. Train a real model on the training split.
5. Evaluate on validation and do error analysis — not just a score, but which cases fail.
6. Iterate on features, data, and model. Most gains come from data, not algorithms.
7. Final check on the untouched test set, once. Then ship and monitor.

Why a baseline first

A baseline is the cheapest insurance in ML:

• It tells you if the problem is even learnable from your data.
• It exposes leakage early (a "too good" baseline is a red flag).
• It sets the reference: a 92%-accurate model is worthless if predicting the majority class gives 91%.
• It's a working end-to-end pipeline you can improve incrementally.

Ship the dumbest model that runs end to end on day one. Optimize from there.

Where time actually goes

Pitfall

Jumping to a complex model before a baseline means you can't tell whether your gains come from the model, a leak, or noise.

Connects to: error analysis · how learning works · pipelines