Predictive AnalyticsActing on likely futures

What predictive analytics actually predicts#

A predictive model looks at patterns in past data and estimates the probability of a future outcome for each customer, lead or campaign. It does not know the future. It ranks likelihoods, and its rankings are only as good as the history it learned from.

• Propensity to buy: how likely a lead or customer is to purchase in a given period.
• Churn risk: how likely a customer is to stop buying or cancel.
• Customer lifetime value: an estimate of future revenue or margin from a customer. See CAC and LTV.
• Next best product or offer: which item a customer is most likely to want next.
• Demand and budget forecasts: expected sales, leads or costs by week, channel or region.

Many marketers already use predictive analytics without building anything. Automated bidding, audience expansion, predicted metrics in analytics tools and recommendation engines are all predictive models inside platforms. Understanding the principles helps you feed them better and question them properly.

From data to decision#

Start with the decision, not the data. 'Which customers should our retention team call this month?' is a decision a churn model can improve. 'Let's see what the data says' usually produces interesting charts and no change in behaviour.

Define the outcome just as precisely. 'Churn' could mean no purchase in 90 days, a cancelled subscription or a downgraded plan, and each produces a different model. Agree the definition with the teams who will act on it, write it down, and do not change it quietly halfway through a quarter.

Finally, decide who will receive the predictions and in what form. A score buried in a database helps nobody. A ranked list in the CRM view a retention manager opens every morning, with the top reasons shown beside each name, gets used.

The data it needs#

Predictive models need enough historical examples of the outcome you care about. If you have had very few churned customers or very few large deals, a model will struggle, and a simple rule written by an experienced person may do better.

• Behavioural data: visits, product views, email engagement, app use, support contacts.
• Transactional data: purchase dates, amounts, categories, returns, payment methods.
• Relationship data: tenure, plan type, account size, channel of acquisition.
• Consent status: which data you are permitted to use for which purpose. See first-party data.

Turning scores into actions#

A churn score on its own is not a strategy. The useful question combines likelihood with value: who is both at risk and worth saving? A simple grid makes the decision visible.

The same logic applies to propensity models. A high propensity to buy might mean the customer will buy anyway, so a discount wastes margin. The most valuable targets are often those whose behaviour your action can change, which is why measurement needs a control group.

Proving it works: the control group#

The only reliable way to know whether acting on predictions helps is to hold back a random group that receives no action and compare outcomes. Without this, you will credit the model for results that would have happened anyway. Our guide to incrementality testing explains the method.

Try changing the churn rate input. Targeting customers at higher predicted risk raises the value of the same action, which is precisely what a good churn model is for.

Where predictive analytics goes wrong#

The predictive models you already feed#

For most marketers, the most important predictive models are not ones they build. They are the bidding and targeting systems inside ad platforms, which predict the likelihood of a conversion for each auction and bid accordingly. These systems learn from the conversion signals you send them.

That makes your conversion tracking a form of model training. If every form fill counts as a conversion, the platform learns to find form fillers, including the ones who never answer the phone. If you send back which leads became qualified opportunities or paying customers, it learns to find more of those. See offline conversion tracking and smart bidding.

• Send the conversion closest to real value that you can measure reliably and in reasonable volume.
• Assign values that reflect real differences between outcomes, where the platform supports it.
• Exclude spam, duplicates and test conversions before they train the system.
• Change conversion definitions deliberately and rarely, since each change resets what the model is learning.

The same principle applies to CRM lead scoring and email send-time features. Predictive tools inside platforms are only as good as the outcomes you define for them. Defining the right outcome is a marketing decision, not a technical one.

Fairness, privacy and explainability#

Predictive models can encode unfair patterns from history, such as offering worse terms to certain postcodes or groups. Review which attributes are used, avoid sensitive characteristics, and check outcomes by segment. Under privacy laws such as India's DPDP Act, use personal data only for purposes customers were told about.

Explainability matters for adoption too. Sales and service teams act on scores they understand. Show the main reasons behind a score, in plain language, alongside the number. For the wider ethical frame see AI ethics in marketing.

Build, buy or use what you have#

Most organisations should exhaust what their existing platforms offer before building. See build vs buy AI marketing tools for the decision in detail.

Published by Fabulous.Media, a network of specialist marketing agencies. Updated 9 October 2026. Platform features change often; check current official documentation before acting on platform-specific detail.