What is a predictive model?
A predictive model is a model built to estimate something that is not yet known from things that are. The unknown may be in the future, such as whether a customer will renew, or hidden in the present, such as whether a transaction is fraudulent. The model learns the relationship from historical records where the outcome was eventually observed, then applies it to new records where it has not been.
Most predictive models are regression or classification models from classic machine learning: logistic regression, gradient-boosted decision trees, and similar methods. They run on structured data from the ERP, CRM, and ticketing systems a company already has. Deep learning is used when the inputs are images, text, or long sequences.
Predictive models in practice
A managed service provider builds a model to predict which client contracts are at risk of not renewing in the next two quarters. The inputs are ticket volume, response times, escalations, invoice disputes, and time since the last account review. Each quarter the model produces a ranked list, and account managers work the top of it. The model does not make the retention decision. It determines which accounts receive a review.
Predictive models sit inside predictive analytics, which is the business practice of using them, and alongside prescriptive analytics, which goes a step further and recommends an action. Their value is realized only when the prediction is connected to a decision that someone owns, and sustained only when model monitoring catches the drift that every predictive model eventually suffers.