"Predictive modelling" sounds like a crystal ball, which is part of why people either overbuy it or dismiss it. The reality is plainer. A predictive model looks at past cases where you know what happened, finds patterns that tend to go with each outcome, and uses them to put a probability or an estimate on new cases.
That's it. The skill is in choosing a question where that helps, and in knowing what to do with the answer.
The questions it suits
Nearly every business use falls into one of three shapes.
How much? Estimating a number: next month's sales, how long a delivery will take, how much a customer will spend over a year. These are regression and forecasting problems.
Which one? Sorting cases into groups: will this customer leave, will this order be cancelled, will this lead buy, is this transaction suspicious. Here the model gives a probability and you decide where to draw the line.
What goes together? Finding structure without a known outcome: groups of customers who behave alike, products that tend to be bought together. This is clustering and association analysis.
If your question doesn't fit one of those, it probably isn't a modelling problem yet. It may be a measurement problem or a decision problem wearing a modelling hat.
What it needs
Three things, and they're not glamorous.
- History with outcomes. You can't learn who churns without records of people who did and didn't. For a customer-level model, a few thousand rows with a known result is a workable start.
- Inputs that were known at the time. This one trips up experienced people. If you're predicting cancellation, you can't use a field that only gets filled in after the cancellation. The model will look brilliant in testing and be useless in real life. It's called leakage and it's the most common reason a promising model falls over.
- A decision at the other end. If nothing changes when the score changes, the model is a hobby.
What "accurate" means
People ask how accurate a model will be as if there were one number. There isn't, for two reasons.
First, accuracy depends on the cost of each kind of mistake. Missing a customer who was about to leave costs you their revenue. Flagging a customer who was never going to leave costs you a discount you didn't need to give. Those aren't equal, and the right threshold depends on the ratio.
Second, the only fair measure is on data the model hasn't seen. You hold some back, or you test on a later period than you trained on, and measure there. Then compare against how you decide today. If a simple rule gets you most of the way, the model has to justify the extra complexity.
The model is the middle of the project
The visible part, picking and tuning an algorithm, is often a small slice of the work. Before it comes the data: finding it, joining it, deciding what each field means. After it comes the part that decides whether anyone benefits: getting the scores somewhere people will see them, agreeing what to do about them, and checking over time that the model still works as customers and products change.
Models go stale. A pattern learned before a price change or a new competitor can quietly stop being true. Someone needs to own the job of checking.
Simple often wins
You'll hear a lot about deep learning and large language models. Most business predictions on tables of customer, order or product data are handled well by far older methods: linear and logistic regression, decision trees, random forests, gradient boosting. They're quicker to build, easier to explain, and usually just as good on this kind of data. We'd rather ship a model your team understands than one that scores a fraction better and nobody dares touch.
A decent first project
Pick one decision that gets made repeatedly, where being a bit better would be worth real money, and where you already keep records of what happened. Write down what you'd do differently with a good prediction. If that sentence comes easily, you've got a candidate.
If you'd like a second opinion on whether a question is ready for modelling, talk to us about predictive modelling. Sometimes the honest answer is "fix the tracking first".