What is predictive analytics?
Predictive analytics is the practice of fitting models to historical data in order to estimate what is likely to happen next. The output is a probability or a forecast attached to a specific entity and horizon: this account has a high chance of paying within 30 days, this compressor is likely to fail within two weeks, demand for this product next month will land in a given range. The methods range from regression and time series forecasting to gradient-boosted trees and neural networks, chosen to fit the data and the question.
It is the third stage in the common progression from descriptive and diagnostic analytics toward prescriptive analytics. The earlier stages establish what happened and why, and predictive work assumes those patterns will continue and projects them forward.
How predictive analytics works
A team assembles a training dataset where each row is one entity at one point in time, with features describing it then and a label recording what happened afterward. A model learns the relationship between features and label, is validated on data it has not seen, and is then scored against current records to produce predictions. Those predictions become useful once a workflow consumes them, such as a work order, a call list, or a reorder quantity. The model is also monitored for drift, because the behavior it learned changes as the business and its customers change.
Predictive analytics in practice
A collections agency builds a propensity-to-pay model from three years of account history, contact attempts, and payment outcomes. Each morning the model scores every open account, and the dialer works accounts in order of expected recovery rather than in order of balance. Collectors spend their time on accounts likely to resolve, and low-probability accounts move to a cheaper letter and text campaign. The agency measures the model against a held-out group of accounts worked the old way, which is the only honest proof that the predictions changed anything.
What to watch for
- Ask for the baseline, because a model that predicts churn is only valuable if it beats the simple rule the team already uses.
- Check for information leakage, where a feature quietly encodes the answer and produces excellent test scores with useless production results.
- Decide the action before building the model, since a prediction nobody can act on in time is just a report.
- In regulated uses such as lending and collections, confirm that the features and outcomes meet fair lending and consumer protection obligations and that the model can be explained to an examiner.
- Plan for retraining from the start, because predictive models decay as conditions change.