Exercise 2: Precision vs. Recall, and Why False Negatives Cost More Here — Possible Solution ==================================================================== PRECISION, PER THIS CHAPTER'S OWN DEFINITION ------------------------------ Per this chapter, "Precision = True Positives ÷ (True Positives + False Positives) — of everyone the model flagged as a flight risk, what fraction actually left?" Precision is a question about the model's OWN positive predictions specifically — out of everyone it raised a flag on, how many of those flags were actually correct. It says nothing about employees the model failed to flag at all. RECALL, PER THIS CHAPTER'S OWN DEFINITION ------------------------------ Per this chapter, "Recall = True Positives ÷ (True Positives + False Negatives) — of everyone who actually left, what fraction did the model catch?" Recall instead starts from the REAL, actual leavers and asks how many of them the model successfully identified. It says nothing about how many false alarms were raised along the way. WHY THESE ARE GENUINELY DIFFERENT QUESTIONS ------------------------------ Precision is anchored to the model's own predictions (how trustworthy is a "yes" from this model?); recall is anchored to the real-world outcomes (how thorough is this model at finding every real case?). A model could have excellent precision (everyone it flags really is at risk) while having terrible recall (it only flags a tiny handful of people, missing most actual leavers) — or the reverse (flagging almost everyone, catching nearly all real leavers, but with a flood of false alarms). The two metrics are answering different questions about different sets of people (the flagged group vs. the actually-left group), which is exactly why both are needed rather than either alone. WHY A FALSE NEGATIVE IS USUALLY MORE COSTLY HERE, PER THIS CHAPTER ------------------------------ Per this chapter, "a missed at-risk employee (a False Negative) is usually the more expensive mistake — a resignation nobody saw coming, with no chance to intervene — while a false alarm (a False Positive) costs, at worst, an unnecessary retention conversation." A False Negative represents a genuine, real-world failure: an employee actually leaves, and because the model never flagged them, nobody had the opportunity to intervene at all — the cost is losing that employee with zero chance to prevent it. A False Positive, by contrast, costs comparatively little: a manager has an unnecessary but generally low-stakes conversation with someone who was never actually planning to leave — an inconvenience, not a lost employee. WHY THIS IS A JUDGMENT CALL, NOT A DATA-DETERMINED FACT ------------------------------ Per this chapter, this is "a genuine, business-specific judgment call about which error type matters more, not a fact the data itself decides." The relative cost of a False Negative vs. a False Positive depends on real business factors (the cost of turnover, the cost of a retention conversation) that live outside the dataset and outside the model itself — the data can report precision and recall numbers, but it can't tell you, on its own, which of the two errors your organization should care about more. WHY THIS WORKS AS AN ANSWER ------------------------------ It defines precision and recall precisely using the chapter's own confusion-matrix-based formulas, explains why they answer genuinely different questions about different groups of people, and explains the chapter's own specific business reasoning for why a False Negative is the costlier error for the attrition problem, while noting this conclusion is a judgment call rather than a mathematical certainty.