Predicting a Score vs. Predicting a Position

Opening Scene

Two coaches are looking at the same stack of player stats, but asking different questions. One wants to know: how many points will this player likely score next game? The answer could be anything — 8, 14.5 on average, 22 on a hot night. The other wants to know: is this player a guard, forward, or center? The answer has to be one of a small, fixed set of options — there’s no “16.3” answer to “what position is this.”

Both are predictions. Both use the same kind of labeled data from Article 1. But they’re asking fundamentally different shapes of question, and a supervised model has to be built differently depending on which shape it’s answering.

In Plain English

Regression predicts a continuous number — points scored, temperature, price. The answer can be almost any value along a range. Classification predicts a category from a fixed, known set of options — guard/forward/center, spam/not-spam, fraud/not-fraud. The answer has to be one of those specific choices, nothing in between. Every supervised learning problem starts by figuring out which of these two shapes it actually is, because it changes what “correct” even means.

The Old Way

Before either term existed formally, people made this same distinction instinctively:

Estimating how long a trip will take (a number, regression-shaped) versus deciding which route to take (a choice among a fixed set of routes, classification-shaped).
Guessing a patient’s temperature versus diagnosing which of a known set of conditions they likely have.
Forecasting how much rain will fall versus deciding whether to call the game rained-out or not.

The intuitive difference was always there — a measurement versus a decision between known options. Formalizing it just gave the distinction a name and a set of matching techniques.

What’s Changing (and Why AI Is the Reason)

The same underlying data increasingly gets modeled both ways for different purposes. A player’s stats can feed a regression model predicting exact points scored and a classification model predicting whether they’ll have a “breakout game” — same input, two different question shapes, both useful.
AI tooling increasingly helps identify which shape a business question actually is before anyone builds anything. Framing a vague ask like “will this customer be a problem” as regression versus classification used to depend on an experienced practitioner’s instinct; that framing step is increasingly assisted rather than purely manual.
Classification is expanding beyond simple categories into far larger option sets — as covered later in this series, some modern classification-shaped problems choose from thousands of possible categories, which strains techniques originally built for a handful of options.

The Metaphor, Fully Extended

Basketball Prediction	Learning Task
Predicting exact points scored next game	Regression — predicting a continuous number
Predicting which position a player plays	Classification — predicting one of a fixed set of categories
A predicted value of 17.5 points	A valid regression output — fractional values make sense
A predicted value of “guard and a half”	A nonsensical classification output — categories don’t blend like numbers do
Being off by 2 points on a scoring prediction	Regression error — measured as a distance from the true number
Predicting “forward” when the player is a “center”	Classification error — a wrong category, not a distance

For Beginners: What to Actually Do

Before building or evaluating any model, ask plainly: is the answer a number along a range, or a choice among a fixed set of options? The rest of the approach follows from that answer.
Get comfortable that “how wrong” means something different in each case — a distance from the true number in regression, right-or-wrong against a category in classification.
Watch for questions that sound like one shape but are secretly the other — “how likely is this customer to churn” is often framed as a probability, which sits closer to classification than it first appears.

For Practitioners and Leaders: The Deeper Layer

Misframing a business question’s shape early — treating a genuinely continuous outcome as a handful of buckets, or vice versa — quietly caps what any model built on top of it can ever achieve.
Some real problems are genuinely both: predicting an exact value and also a confident range around it blends regression thinking with classification-style confidence, and deserves to be evaluated as such.
When a stakeholder asks for a prediction, get the target’s shape confirmed explicitly before any modeling work starts — it’s one of the cheapest questions to ask and one of the most expensive to get wrong late.

Quick Recap

Regression predicts a continuous number; classification predicts a category from a fixed set of options.
The distinction is old and intuitive — a measurement versus a decision among known choices — formalized with matching techniques.
The same data can support both kinds of models for different purposes.
Confirming which shape a question actually is should happen before any modeling begins, not after.

Where This Fits in the Series

Article 3 examined what a label actually is; this article covered the two basic shapes a label can take. Article 5 moves into unsupervised territory again, looking at how players naturally group by how they move — the beginning of clustering.