A pricing actuary says "let's predict lapses." Before any algorithm, you check three things: is there a measured outcome, are there usable predictors, and is the goal a prediction on new records. Those three define a predictive modeling problem.
A measured target. Every predictive problem centers on one outcome variable you want to forecast: claim severity, lapse yes or no, next-year revenue. If you cannot point to the target as a column of data, you do not yet have a predictive problem.
Supervised by design. Predictive modeling is supervised learning. The training data must carry the observed target on each historical record, so the algorithm can learn how predictors map to outcome. Unsupervised methods like clustering and principal components have no target and answer a different question.
KEY: The dividing line is the label. If each historical record carries the outcome you want to predict, the problem is supervised and predictive. If not, you must collect that outcome or reframe the task as unsupervised.
Common mistakes
- Calling clustering predictive. No target column means unsupervised. A predictive problem must name the outcome you forecast.
- Trusting accuracy on a rare target. With a 1.2% base rate, always predicting "no" scores 98.8% yet catches nothing. Check the base rate first.
- Using a leaked predictor. A field recorded at or after the event, like final claim status, cannot be an input to predict that same event.
Bottom line
- A predictive modeling problem centers on one target variable, the outcome you want to predict.
- It is supervised: the historical data records the observed target on every row.
- Predictors (features) are inputs that plausibly relate to the target and are known before the outcome.
- The goal is generalization: accuracy on new, unseen data, not on the training set.
Exam shortcut
First locate the target column: present and observed means supervised and predictive; absent means unsupervised or a data-collection task. Read the target's type next, numeric points to regression and categorical points to classification. For any rare target, distrust accuracy and compute the base rate before trusting any performance metric.
The full lesson (about 1,121 words, 7 min read) adds 2 worked examples, all 6 common mistakes, a self-check, free in the app.
Learning objectives
- 1b
Browse all free Exam PA lessons or jump into free Exam PA practice questions.