What is predictive analytics

Predictive analytics defined

Harvard Business School describes predictive analytics as using data to predict future trends and events. That is the core predictive analytics meaning.

The definition of predictive analytics has three parts: historical data, a model that finds patterns, and a forecast about something you care about.

Teams use the output to act earlier, for example by staffing up before a busy week or calling a customer who might leave.

To define predictive analytics in one line: past data used to estimate what is likely to happen next.

How predictive analytics works

Most projects follow the same path. You pick a question, prepare the data, train a model and test it on records it has not seen.

One common tool is regression analysis, which measures the relationship between two variables, or between three or more in multiple regression, per Harvard Business School.

Testing matters. Forecasting research describes rolling origin evaluation, where the forecast start point moves through the time series so the model is tested over several periods, not one lucky stretch.

Healthcare and marketing examples

• Demand

  Estimate orders before they arrive.

• Churn

  Spot customers who may cancel.

• Risk

  Flag claims or loans for a closer look.

• Healthcare

  Find patients who may need follow-up.

• Marketing

  Score leads that are ready to buy.

Healthcare predictive analytics estimates which patients may miss appointments, return to the hospital or need extra follow-up. Read more in our AI in healthcare guide.

Predictive analytics and marketing go together too. Marketers score which leads are likely to buy and which customers are about to churn. Predictive analytics for marketers also helps time campaigns.

Operations teams forecast demand, stock and cash flow. Our AI forecasting work covers this kind of estimate.

Health care predictive analytics follows the same steps with patient records, under stricter privacy rules.

Limits and data needs

A prediction is only as good as the data behind it. Missing, messy or outdated records weaken the result.

Accuracy alone can hide a problem. Researchers note a model can sit close to actual values yet overestimate more often than it underestimates, which can hurt a business.

Track the direction of errors, not just their size.

NIST also advises documenting where a model was built and where it may not hold up.

Results in practice are modest: a Stanford survey summary reports that among companies using AI in marketing and sales, 71% see revenue gains, but the most common gain is under 5%.