Day 1 · Monday

From a question to a prediction you can trust

Before you trust a prediction, you need to know exactly what it was asked to do. Today we build that standard piece by piece, and then we apply it to a real claim.

By the end of today

• frame a business question as a prediction task with a written contract;
• name the loss and the baseline that judge a prediction;
• explain why we split data into training, validation, and test sets;
• spot a feature that leaks information from after the prediction moment;
• say what a 90% accuracy figure does and does not show.

What we cover

• slides 5–11Part I: What machine learning isShort. Where these methods sit and what they estimate.
• slides 12–23Part II: The prediction contractWhat must be written down before anyone fits anything.
• slides 24–34Part III: Learning from dataLoss, baselines, and the three-way split.
• slides 35–51Part IV: The limitsLeakage, decayed models, and what a score can and cannot support.

Slides 52–57 are optional appendix material for students who want the full algebra.

Lecture materials

Review questions

• What baseline did the 90% model have to beat, and what was its value?
• A new column makes your validation score jump. What do you check first?
• Why may the test set be used only once?
• Accuracy 90.2% and recall 23.3% describe the same model. Explain how both can be true.
• Why does a high response probability alone not show that calling causes subscriptions?

← All four days  ·  Day 2 →