Prediction: your first ML idea

You built the ML loop in miniature: predict → score → adjust. You didn’t train a deep net yet — you felt the idea that “learning” means your rule gets better on measured examples. Point at the accuracy number: “I changed the threshold; that went up.”

A tiny predictor is often a threshold on a score:

truth = [1, 0, 1, 1, 0]
scores = [0.8, 0.3, 0.6, 0.9, 0.2]
threshold = 0.5

preds = [1 if s >= threshold else 0 for s in scores]
accuracy = sum(p == t for p, t in zip(preds, truth)) / len(truth)

1. Features / scores — what the model “sees”
2. Label — the correct answer for each row
3. Rule — e.g. “score ≥ threshold → yes”
4. Accuracy — fraction of matches

Mental model: ML is practicing on past quizzes, then taking a new quiz.

• Deciding spam vs not from a spamminess score
• Approving a loan only above a risk cutoff
• Tuning “how sure” you need to be before auto-acting

High accuracy on a tiny toy set can fool you. One lucky threshold that memorizes ten examples may fail on the eleventh. Always ask: “Would this hold on new rows?”

Slide the threshold higher, then lower. Note one setting where accuracy drops. That’s evidence you’re really steering the rule.