Predictive Analytics: What It Is and How to Apply It

A subscription-based clinic closes the month and finds out, on the 2nd, that it lost 38 customers in March. Everyone looks at the number, sighs, and moves on to the next slide. The uncomfortable part sits elsewhere: in nearly all 38 cases, the warning signs had been sitting in the database for weeks. A session cancelled and never rebooked, visit frequency dropping, two invoices paid late.

Nobody there needed a crystal ball. They needed someone sweeping the database every day looking for that pattern, which no team does by hand across a thousand customers spread over three systems.

That is the gap predictive analytics fills. And unlike five years ago, most of it no longer requires hiring a data scientist.

What is predictive analytics?

Predictive analytics is the use of historical data to estimate the probability of a future event. An algorithm learns which combinations of behaviour came before an outcome (churn, purchase, default) and applies that pattern to the current base, returning a probability score per customer, product or period. It is not fortune telling. It is counting patterns that already repeated.

The word that matters is probability. An honest model never says "this customer will churn". It says "customers behaving like this churned 82% of the time over the last two years". Whether you call, discount or do nothing is still your call, and that is where most of the value is won or lost.

Predictive, descriptive and prescriptive: what is the difference?

All four categories show up in every BI deck, usually scrambled. The split is simple and worth memorising, because it defines the kind of question each one answers.

Type	Question it answers	Retail example	What it requires
Descriptive	What happened?	Sales dropped 12% in August	Organised history
Diagnostic	Why did it happen?	Category X lost paid traffic	Data joined across channels
Predictive	What is likely next?	Category X should fall another 8%	History plus explanatory variables
Prescriptive	What do I do now?	Move 6k in budget to category Y	Business rules and real constraints

In practice, most companies still live between the first two rows. Jumping straight to prediction before solving those two usually produces a pretty model with bad data inside it. If you cannot yet explain why sales fell, start with a structured diagnosis before thinking about forecasting.

How a predictive model actually works

Every predictive model repeats four stages, whatever the tool. First you define the event you want to predict, and it has to be binary and dated: churned or not, bought or not, within the next 30 days. "Increase engagement" is not a predictable target, it is a wish.

Then comes history. The algorithm needs to see the event happen many times, together with the behaviour that preceded it. Only then does it compute weights: falling usage may matter more than a late payment, or the other way around, depending on your business. The output is a score per record.

The fourth stage is the one almost nobody plans: what happens to the list. A report naming 400 high-risk customers sitting in a shared folder changes nothing. It has to become a CRM task, a call queue, an email flow. A model with no route to action is pure cost.

How much history do you need?

It depends on how rare the event is, not on how big your company is. The rule of thumb: the algorithm needs a few hundred positive examples to find a pattern, and a similar volume of negative examples to know what to compare against.

One public benchmark helps calibrate expectations. To unlock its native predictive metrics, Google Analytics 4 documentation requires at least 1,000 returning users who triggered the condition (purchase or churn) and 1,000 who did not, over a seven-day window within the last 28 days. That threshold is not arbitrary. It is roughly the volume any model needs to avoid mistaking coincidence for pattern.

If your base is smaller, per-customer prediction is not the right move yet. Aggregate forecasting, such as monthly revenue or demand per category, remains viable with far less data.

The three use cases that pay off fastest

Churn risk

The classic case, because the return is easy to measure. Here is a simulated example with round numbers to show the logic: a base of 4,000 subscribers with a monthly churn rate of 4% loses 160 customers a month. A decent model tends to concentrate a good share of those losses in the highest-risk group. Say the riskiest 10%, or 400 accounts, account for 60% of cancellations, which is 96 customers.

If retention efforts reverse a quarter of those, that is 24 accounts saved per month. At an average ticket of US$ 40, that preserves US$ 960 in recurring revenue every month, with the team talking to 400 people instead of 4,000. The gain did not come from the algorithm. It came from no longer treating the whole base the same way.

Demand forecasting

Anyone carrying inventory feels this one first. Demand forecasting uses seasonality, trend and known events (holidays, campaigns, retail dates) to estimate how much of each item will move. The expensive mistake is rarely the stockout on your best seller, which everyone notices the same day. It is the capital frozen in mid-velocity items that nobody tracks one by one. Worth reading as well: how AI sales forecasting works in practice.

Purchase propensity

Here the model ranks your base by probability of buying a given offer. It decides who sales calls first, who gets the coupon and, above all, stops you discounting people who would have bought anyway. That last point usually carries the highest return and gets the least attention.

How to apply predictive analytics without coding

There are three levels, and most companies should start with the first.

1. Built-in features. GA4, ecommerce platforms, CRMs and ERPs already ship predictions: purchase probability, churn risk, revenue projection. Zero cost, zero control. Good enough to test whether a forecast changes any decision before you invest.
2. AI on your own data. Conversational analytics tools connect the sources you already have and answer questions in plain language, projections included. This is where most mid-sized companies solve the problem today, with no IT project.
3. Custom model. A team trains and maintains its own model for one specific case. It makes sense when the decision repeats, carries high value, and the two levels above already proved the return.

The common mistake is starting at level three. Six months and a serious budget later, you find out nobody in operations uses the list the model produces. If automating decisions interests you, first understand what AI agents are and what they actually solve.

Some context before you feel behind: according to the TIC Empresas 2025 survey by Cetic.br, 17% of Brazilian companies used some form of artificial intelligence in 2025. Among small companies it was 15%; among large ones, 50%. The gap between those two groups is exactly where the opportunity sits for whoever is smaller and decides faster.

Where predictive analytics breaks

High accuracy can be useless. If your monthly churn is 4%, a model that answers "nobody will churn" for everyone is right 96% of the time and useless. Always ask how many real cancellations the model captured, not what its overall hit rate is.

Models learn the past. Changed your pricing, your mix, your shipping rules or the ad platform algorithm? The learned pattern just aged. A demand forecast built before a major commercial policy change is worth little.

Dirty data comes out looking like certainty. Duplicate customers, dates typed by hand, statuses nobody updates in the CRM. The model will not warn you that the input is bad, it just returns a score with the appearance of precision. Worth checking the silent data quality errors before trusting any forecast.

The forecast interferes with the outcome. If you gave 30% off to everyone flagged as high risk, churn in that group falls and the model looks wrong. Hold back a control group with no action, or you will never know whether the prediction worked.

Frequently asked questions

What is the difference between predictive and prescriptive analytics?

Predictive estimates what is likely to happen and returns a probability. Prescriptive takes that probability, adds business rules and constraints, and recommends the action. One answers "what is the risk"; the other, "what should I do about it given the budget I have".

Is predictive analytics the same as machine learning?

No. Machine learning is one of the techniques used to do predictive analytics, and not always the best one. A simple regression or a time series model solves plenty of demand forecasting cases, with the advantage of being explainable to whoever has to decide.

Do I need big data to start?

No. You need consistent, dated data, which is different from abundant data. A base of 3,000 customers with two clean years of history yields more useful forecasting than millions of unstandardised records.

How do I know the model is any good?

Compare it against the dumbest alternative. If forecasting with last month's average beats the model, the model is not worth running. For classification problems, look at how many real cases it captured inside the highest-risk group, not at overall accuracy.

Does predictive analytics work for small businesses?

It does, at a smaller scope. Instead of a per-customer model, start with aggregate forecasting: monthly revenue, demand per category, weekly seasonality. Those need little history and already change purchasing and staffing decisions.

Where to start on Monday

Pick a decision you make every week that currently runs on gut feel: how much to buy, who to call, where to move budget. Pull the history of that decision, check whether the outcome was recorded anywhere, and count how many cases you have. A few hundred means a forecast is possible. Fewer than that, and this week's job is to start recording, not to buy a tool.

Sherlok covers the second step of that ladder: it connects Meta Ads, Google Ads, GA4, CRMs, spreadsheets and SQL databases and answers, in plain language, what your own data suggests for the months ahead, with no IT project and no code. If your problem today is reaching the question before the month closes, take a look at Sherlok.