Predictive analytics

What is predictive analytics?

Predictive analytics uses historical data to estimate what is likely to happen next. A normal dashboard tells you what happened last month. A predictive model estimates next month's demand, which customers may churn, which invoices may be paid late, or which machine may fail soon.

The output is usually a probability, score, category, or expected value. It is not a guarantee. It is a structured way to act earlier than you could with reporting alone.

The weather forecast is a useful analogy. Past measurements and current conditions do not make tomorrow certain, but they give you enough signal to take an umbrella.

Predictive analytics in the analytics ladder

Analytics is often described in four levels.

Descriptive analytics. What happened? Revenue last month, open tickets, stock on hand.

Diagnostic analytics. Why did it happen? Which product, channel, customer group, or process caused the change?

Predictive analytics. What is likely to happen? Demand, churn, risk, failure, delay, or conversion.

Prescriptive analytics. What should we do? Recommended order quantity, next-best action, route change, price change, or intervention.

You do not need a perfect analytics maturity model before predicting. You do need the same foundation: reliable history, clear definitions, and a way to compare predictions with what actually happened.

How a prediction model works

Most predictive analytics starts with examples from the past. For churn prediction, each row might describe a customer at a point in time: usage, complaints, contract age, support tickets, payment behaviour, and whether that customer later left.

The model learns which patterns are associated with the outcome. A classification model predicts a category or probability, such as churn risk. A regression model predicts a number, such as next month's units sold. A time-series model forecasts values over time, such as weekly demand.

Before using the model, you test it on data it did not see during training. After launch, you keep comparing predictions with real outcomes. That monitoring is what tells you whether the model still works.

Where predictive analytics helps

• Demand forecasting. Estimate sales by product, store, channel, or week so stock and staffing can be planned earlier.

• Churn prediction. Score customers by risk of leaving, so the team can focus retention work where it has a chance to matter.

• Late-payment risk. Flag invoices or customers that are more likely to pay late, and adapt follow-up.

• Predictive maintenance. Use sensor history and breakdown records to plan service before a machine stops.

• Lead scoring. Estimate which leads are more likely to convert based on source, behaviour, company profile, and previous sales outcomes.

The common thread: acting earlier has economic value, and the data contains repeated patterns you can learn from.

Predictive analytics versus reporting and forecasting

A report or dashboard is still the right tool when you need shared visibility into what happened. Predictive analytics is worth the extra work when a timely action can change the outcome: order stock before the shortage, call a customer before they leave, maintain a machine before downtime.

Forecasting is a narrower form of prediction focused on time series. It projects a value forward through time, often using trend, seasonality, and recent behaviour. Predictive analytics is broader. A churn model predicts an outcome per customer, and a fraud model predicts risk per transaction.

Predictive analytics also differs from generative AI. A prediction model estimates a probability or value from data. A generative model creates new content such as text, images, code, or audio. They can be combined, but they solve different problems.

What you need before starting

History with outcomes. The event you want to predict must appear often enough in the past: churned customers, late invoices, failed machines, returned products, converted leads.

Data in one place. Useful signals often live across ERP, CRM, support, ecommerce, marketing, and sensor systems. A data warehouse or lakehouse is often the hard part of the project.

A clear action. A prediction that nobody acts on is decoration. Decide who receives the score and what they do with it.

Feedback after launch. Capture the real outcome so the model can be evaluated and retrained.

What to watch out for with predictive analytics

Prediction is not certainty. A customer with high churn risk can still stay. Treat scores as decision support, especially when people are affected.

Bias can be learned from history. If past decisions were unfair or incomplete, the model may repeat that pattern.

Model drift is normal. Prices, products, customer behaviour, seasonality, and processes change. Track performance and retrain when the model gets stale.

Data leakage can fool you. Do not train on fields that would not be known at prediction time. For example, a cancellation reason recorded after churn cannot be used to predict churn before it happens.

Start with one decision. A narrow use case with a clear action beats a broad prediction platform with no owner.