Predictive analytics
Predictive analytics is the use of historical data, statistical models and machine learning to estimate what is likely to happen next, such as which customers will churn, which leads will convert or how demand will move.
How it works
A predictive project starts with a question and an outcome to predict, then gathers historical data in which that outcome is known. Analysts choose features, the variables that may predict the outcome, and train models such as logistic regression, decision trees, random forests or neural networks. The model is tested on data it has not seen, and its accuracy decides whether it is used.
Predictions come as probabilities or scores: a customer with a churn score of 0.8 is more likely to leave than one at 0.2. Galit Shmueli's paper in Statistical Science separates models built to predict from models built to explain, and shows that the two goals call for different choices of variables and methods, so the model that forecasts churn best is not necessarily the one that identifies its causes.
A worked example
Of 900,000 subscribers, a streaming service's churn model flags 12,000 as high risk, with scores above 0.7. A retention offer to half of them, 6,000 subscribers, cuts churn in that group to 18 percent, against 29 percent in the half that got no offer.
Related
Automate this in Pulse
A churn model's top-ranked features are its guess at the cause, and interviewing customers who left through Pulse tests that guess against the reasons people give.