Predictive Analysis

Estimating what will happen, and saying how sure you are.

Also known as: predictive analysis, predictive modelling, prediction

Predictive analysis estimates future or unseen outcomes from past data: which leads will convert, which customers will churn, next month’s demand. It differs from forecasting mainly in scope — forecasting extrapolates series over time, while prediction scores cases (users, transactions, days) from their features.

features (past behaviour, attributes) ──▶ model ──▶ score per case + validation

The credibility of a prediction rests on validation, not sophistication. A model tested on data it never saw, scored with the metric the decision actually needs, and compared against a simple baseline (naive for series, majority-class or a rule of thumb for cases) is trustworthy in proportion to that evidence. Everything else is a demo.

The classic mistakes:

No baseline. A churn model at 85% accuracy where “predict no churn for everyone” scores 90% is worse than nothing. Baselines first, always.
Leakage. A feature that contains the answer — cancellation date predicting churn, future information in a “historical” snapshot. Leakage inflates validation and fails live; audit features for time-travel.
Wrong metric. Accuracy on imbalanced classes, MAPE on intermittent demand. Score with the loss the decision cares about (see classification vs regression).
Overfitting the validation. Tuning repeatedly against one holdout turns it into training data. Lock a final test set, or accept that the reported number is optimistic.
Shipping scores without a decision rule. A churn probability nobody acts on is decoration. Define the threshold, the action and the owner before deploying.

When not to predict: when a rule suffices (“trial ended yesterday” needs no model), when the data cannot support the claim, or when the prediction would not change any action. Prediction is a means to a decision — start from the decision and work backwards (see prescriptive analysis).