A hospital executive says "our readmissions are a problem." That sentence cannot be modeled. Your job is to convert it into a target variable, a unit of observation, and predictors drawn from real data fields.
From vague to analyzable. A business owner hands you a concern, never a model spec. "Costs are too high" or "customers keep leaving" describes a symptom. You must translate it. The translation has three moving parts.
The target variable. Name the single outcome you will predict. It must be measurable and present (or collectable) in the data. "Reduce churn" becomes "predict whether a policyholder cancels within 12 months." That target is a yes/no label, so the task is classification. "Costs are high" becomes "predict claim severity in dollars," a numeric target, so the task is regression.
KEY: Read the target to pick the model family. A category (cancels or not, high/medium/low) is classification. A number (dollars, count, days) is regression. Stating the target first settles the method.
Common mistakes
- Leaving the target vague. "Predict customer behavior" is not a target. Name the exact measurable outcome, such as account closure within six months.
- Skipping the unit of observation. Failing to say "per patient" or "per policy" makes the prediction meaningless. Fix the row definition first.
- Ignoring the data dictionary. Predictors described in the abstract lose credit. Cite actual fields like balance or complaint count.
Bottom line
- A vague business question is not analyzable until you name a measurable target variable.
- Define the unit of observation (per patient, per policy, per claim) explicitly.
- List candidate predictors by referencing actual data-dictionary fields, not vague concepts.
- Classification predicts a category; regression predicts a number. Pick one from the target.
Exam shortcut
Translate in three moves: name the target variable, state the unit of observation, then list predictors tied to real fields. Read the target's type to fix the model family, category means classification, number means regression. Before promising a model, confirm the historical outcome label exists; if it does not, write "collect the outcome data first" as the next step, which graders reward.
The full lesson (about 1,189 words, 8 min read) adds 2 worked examples, all 6 common mistakes, a self-check, free in the app.
Learning objectives
- 1d
Browse all free Exam PA lessons or jump into free Exam PA practice questions.