What is predictive analytics?

Predictive analytics is forecasting a future outcome from patterns in historical data. A model learns from what happened before, for example which customers cancelled or which deals closed, and returns a probability or an expected value for a new situation. The result is not a certainty but a likelihood, and its quality stands or falls with the data the model was trained on.

How does predictive analytics differ from reporting?

Analytics is often divided into four layers. It helps to see where predictive analytics sits:

Most companies' reporting sits in the first two layers. A dashboard in Power BI or an export from Exact, Xero or NetSuite shows what has happened. Predictive analytics adds a step: it calculates what will happen if the pattern continues. That is most valuable when there is still time to act.

How does a predictive model work?

Under the bonnet, most predictive models are surprisingly down to earth. The process always has the same four steps.

1. Choose an outcome. You decide what you want to predict. Will a customer cancel within twelve months? Will this deal close this quarter? Will this invoice be paid late? The sharper the outcome, the more useful the model.

2. Collect historical examples. The model needs cases where you already know the outcome. For example, every customer from the past three years, with for each one whether they left or stayed. More examples are better, but above all, more representative examples are better.

3. Define the features. For each example you collect the attributes that might say something about the outcome. Think of order frequency, change in order size, number of support tickets, payment behaviour, time since the last contact with the account manager, contract length. This is called feature engineering and it is usually the bulk of the work.

4. Train and test. The model looks for the relationship between features and outcome. You then test it on cases it has not seen. Only if it predicts well there is it worth anything.

Techniques range from regression and decision trees to gradient boosting and neural networks. For structured business data, meaning tables of customers, orders and invoices, the simpler methods are often at least as good as the complex ones. They are also easier to explain, and in a financial context that is not a detail.

Where does predictive analytics affect revenue?

In a B2B company there are four applications that bear directly on revenue.

Predicting churn and contraction

A customer who cancels shows up in your figures. A customer who is quietly buying less does not. A predictive model can recognise the pattern that precedes departure: longer gaps between orders, smaller orders, fewer users on a licence, a rise in complaints. This ties in with how you spot customers who are quietly buying less, but looking forward.

Deals and forecasting

A CRM gives each stage a fixed probability: proposal sent is 50 percent, negotiation is 75 percent. A predictive model looks at the attributes of the deal itself: how long has it been stalled, how often has the close date moved, has there been contact with a decision-maker. How that changes a forecast is explained in AI revenue forecasting explained.

Payment behaviour

Which invoices are likely to be paid late? A customer's payment history, the invoice size and the time of year often predict this reasonably well. That makes it possible to send reminders earlier to the customers where it matters, instead of sending everyone the same reminder.

Likelihood of leakage

Less well known, but the most interesting for revenue control: which contracts, orders or customers have an elevated chance of an invoicing error? A contract with an unusual indexation clause, several pricing arrangements and a manual invoicing step carries more risk than a standard subscription. That is predictive analytics applied to the question of whether AI can predict revenue leakage.

Worked example: what a churn prediction is worth

Worked example: suppose you have 400 recurring customers with average annual revenue of EUR 25,000 each. Every year 40 of them leave. That is EUR 1,000,000 of annual revenue walking out of the door.

Each quarter a model flags the 30 customers with the highest risk of leaving. Suppose that on average 12 of them are genuine leavers, and that your account managers manage to keep 4 of those through early contact.

• Retained per quarter: 4 customers x EUR 25,000 = EUR 100,000 of annual revenue.
• Over a year: EUR 400,000 of annual revenue that would otherwise have left.
• Cost: account managers' time for 120 extra conversations a year.

These figures are assumptions to show the mechanism. The two numbers that decide everything in practice are how many genuine leavers the model finds and how many of those your team wins back. The first you can measure by testing the model on historical data. The second only by doing it.

Where does predictive analytics go wrong?

Predictive analytics rarely fails because of the algorithm. It fails because of the data and because of how the output is used.

Poor historical data. A model learns from the past. If your CRM is full of deals that were never closed, close dates nobody updates and customers that appear twice, the model learns that pattern. That is exactly why CRM data is not the same as financial data.

Too few examples. A company with 60 customers and 5 leavers a year has too few cases to train a reliable churn model. Simple rules, such as "no order in 90 days when the average gap is 30", are then more honest than a model that gives false precision.

Leakage inside the model. A classic mistake: using a feature that is only known after the outcome has already happened. For example, "cancellation reason filled in" as a feature to predict cancellation. The model scores brilliantly in testing and is worthless in practice.

Changing circumstances. A model trained on three quiet years predicts poorly in a year with a price increase, a new product or a recession. Models have to be re-tested, not built once and forgotten.

A score without action. A list of at-risk customers sitting in a dashboard that nobody acts on delivers nothing. The value lies in what happens next: who calls, when, with what offer.

How do you judge whether a prediction is any good?

You do not need to be a data scientist to take a critical look at a predictive model. You can put these questions to any vendor or internal analyst:

1. What exactly does it predict? A sharp outcome with a time horizon. "Customer health" is not an outcome. "Cancellation within twelve months" is.
2. What data was it trained on, and over what period? And is that period representative of now?
3. How was it tested? On data the model had not seen before, preferably from a later period than the training data.
4. How often is it right, and how often wrong? Ask for both: how many genuine cases it finds and how many false alarms it raises. A model that flags every customer as at risk finds every leaver and is useless.
5. Why does it give this score? A good model shows, for each prediction, which features weighed most heavily. Without that explanation you cannot check a prediction.
6. What happens to the output? Who gets the list, and what do they do with it?

How does predictive analytics relate to AI?

Predictive analytics is a form of AI, but a different one from the language models that get most of the attention today. A predictive model returns a number: a probability, an amount, a date. A language model returns text. The difference is large and often overlooked; it is worked out in predictive AI vs generative AI.

In a revenue intelligence system the two often work together. A predictive model calculates which customers are at risk or which contracts deviate. A language model explains the finding in plain language or proposes a next step. How these layers fit together is covered in how does AI work within Revenue Intelligence.

How do you start small?

1. Choose one outcome that directly affects money, for example customer contraction or late payment.
2. Check that you have at least two to three years of reliable historical data on that outcome, from the leading system (usually the accounting system, not the CRM).
3. Start with a simple rule and measure how well it predicts. That is your baseline.
4. Only build or buy a model if it is demonstrably better than that rule.
5. Agree who receives the output and what they do with it, before the model goes live.
6. Check every quarter whether the predictions still match what actually happened.

Frequently asked questions

Is predictive analytics the same as forecasting?

Forecasting is one application of predictive analytics. A revenue forecast predicts a total amount over a period. Predictive analytics is broader: it can also predict which customer will leave, which invoice will be paid late or which contract contains an error.

How much data do you need?

There is no fixed number. It depends on how often the outcome occurs. For a rare outcome, such as the loss of a large customer, you need more history than for a common one, such as a late payment. If you only have a handful of cases a year, simple rules are often more honest.

Do you need a data scientist?

To build models yourself, yes. To use and assess them, no. The questions in the checklist above are enough to see whether a prediction is well founded.

Can predictive analytics find invoicing errors?

It can predict where errors are likely, so you check there first. Actually finding an error, such as a missed indexation, is done by comparison: contract against invoice. Predicting and comparing complement each other.