Descriptive vs. Diagnostic vs. Predictive Analysis: Which One Your Question Needs

Every “intro to data analysis” article gives you the same four flashcards: descriptive, diagnostic, predictive, prescriptive. You memorize them, you nod, and then a real question lands on your desk and the list is no help at all — because it never tells you which one you’re actually being asked for.

Here’s the shortcut nobody frames it as: the type of analysis is decided by the verb in the question. Get the verb right and the method, the tools, and the amount of work all fall out of it. Get it wrong and you either over-build (a forecasting model for what was really a one-line lookup) or under-deliver (a number when someone wanted a reason).

1. Descriptive — “what happened?”

The verb is was, did, or how many. You’re summarizing what already occurred. “What was revenue last quarter?” “How many users churned in March?” “What’s the average order value by region?”

This is 80% of the analysis anyone actually asks for, and it’s mostly counting, summing, grouping, and averaging. It feels too simple to call “analysis,” which is exactly why people skip it and regret it — every other type is built on top of a solid descriptive layer. If you can’t cleanly say what happened, you have no business saying why.

2. Diagnostic — “why did it happen?”

The verb is why. Churn jumped in March — why? This is where you slice, compare segments, and look for correlations: did churn rise across the board, or only in one plan tier, one acquisition channel, one region? Diagnostic work is detective work, and it’s mostly about comparison — this group vs. that group, this period vs. the last.

The trap here is mistaking correlation for cause. Diagnostic analysis can tell you churned users were three times more likely to have hit a billing error; it cannot tell you the billing error caused the churn. It narrows the suspect list. Proving cause usually needs an experiment, which is a different job.

3. Predictive — “what will happen?”

The verb is will. “Which customers are likely to churn next month?” “What will demand look like in Q4?” Now you’re fitting a model to historical patterns and projecting them forward.

Two honest cautions, because this is the type everyone wants to jump straight to. First, prediction is only as good as the descriptive and diagnostic work under it — a model fit on data you don’t understand will confidently predict garbage. Second, it’s hungry: it needs enough history, enough examples of the thing you’re predicting, and the assumption that the future resembles the past. If you have 40 rows and three churned customers, you don’t need a model, you need more data.

4. Prescriptive — “what should we do?”

The verb is should. This sits on top of a prediction and adds a recommendation: given who’s likely to churn, which of them should we spend a retention offer on, and how much? In practice this is the rarest type and the one most often faked — a lot of “prescriptive analytics” is a predictive model with a human making the actual call. That’s fine. The judgment is usually the point.

The heuristic, in one line

Read the question and find the verb:

• was / how many → descriptive
• why → diagnostic
• will → predictive
• should → prescriptive

And one ordering rule that saves more time than any tool: you can’t skip levels. Most failed analyses are someone demanding a prediction (“who’s going to churn?”) before anyone has done the descriptive work (“wait — how are we even defining churn?”). When a request feels too hard, it’s usually because it skipped a level. Drop back down: describe it, then diagnose it, and the harder question often answers itself — or turns out to be the wrong question.

A worked example

Say sales dipped last month and your boss wants to know “what’s going on.” That phrase hides all four types. Unpacked, it’s:

• How much did sales drop, and where? (descriptive)
• Why — which segment, product, or channel drove it? (diagnostic)
• Is it likely to continue? (predictive)
• What should we do about it? (prescriptive)

Answer them in that order and each one is small and tractable. Try to leap to “what should we do” first and you’ll be guessing. The whole skill is hearing a vague question and sorting it into these buckets before you touch the data.

The point

The four types aren’t a syllabus — they’re a sorting hat for questions. Most of the value is in the first two, most of the mistakes come from skipping to the last two, and the verb tells you which one you’re in.

If you want to work through the descriptive and diagnostic layers quickly, the fastest path is to just ask the question in plain language and read the result — which is what we built Curator to do. The no-signup summary statistics tool is a decent place to start the descriptive layer on a file you already have.