Predictive Analytics

Short answer: Predictive analytics produces probabilities, scores or forecasts. Examples are an 80% chance that a customer renews, or expected sales of 5,000 units next month. Deciding what to do about that estimate is a separate step. People often make that decision, sometimes with help from prescriptive analytics, which recommends an action.

Common business uses include demand and sales forecasting, predicting which customers will cancel, spotting fraud, scoring sales leads and predicting equipment failures. Methods range from traditional statistics to machine learning models that handle many factors at once. One traditional method estimates how one measure moves with another, such as sales with price. Another projects a series of past figures, such as weekly sales, forward in time. Every prediction carries uncertainty. A well-built forecast states a likely range as well as a single figure. A predictive model is only useful if it beats a simple alternative, such as repeating last year's figure for the same month.

The team defines exactly what to predict and over what period, such as whether a patient will miss next week's appointment. It gathers past records where the outcome is already known. It adds information that was available at the time, such as booking history, appointment type and how far ahead the patient booked. The team trains the model on part of the history and tests it on a separate part the model has not seen. It compares the model's accuracy with a simple method, such as the clinic's average no-show rate. If the model does well enough, the team connects it to live systems so it scores new bookings automatically. Someone then tracks its accuracy, and the team retrains it as conditions change.

A clinic group books 1,000 appointments a month. A model flags 150 of them as likely no-shows. In tests on past data, patients missed 60% of flagged appointments, against 10% of the others. That suggests about 90 missed appointments among the 150 flagged and 85 among the other 850, or 175 in total. The front desk calls only the 150 flagged patients the day before, which is 15% of bookings. Those calls reach about 90 of the 175 likely no-shows, roughly half. Any slots that patients cancel go to people on the waiting list.

Predictions help leaders plan stock, staff and sales effort before outcomes arrive. Their value comes from changing an action, such as ordering differently or calling certain customers first. Before a predictive project starts, the person asking for it should name three things. These are the decision it will change, how the team will measure accuracy and what a wrong prediction costs.

Models assume the future will look like the past. So they can fail when markets, prices or customer behavior shift. They can repeat bias in the historical data. A model that fits past data very closely may do poorly on new data. Precise-looking scores can hide wide uncertainty. Acting on a prediction can change the outcome, which makes later accuracy harder to judge. For example, a reminder call may prevent the very no-show that the model predicted.