Predictive analytics, propensity and personalisation

What does predictive analytics predict?

Predictive analytics is one branch of AI in marketing: it estimates a defined outcome from information available at the time a decision is made. Its usefulness depends on that definition as much as on the model.

“Customer value” is not a sufficiently precise target. The team needs to specify whether it means revenue, contribution, repeat purchase or another outcome, and over which period. A model trained on gross orders may favour customers who later return purchases. A model trained on clicks may identify curiosity rather than commercial intent.

A propensity score is usually an estimated probability or a ranking derived from one. A personalisation system then chooses an action: a reminder, a product recommendation, a service intervention or no contact. These are separate layers. A good score can still feed a poor action.

Consider a hypothetical Indian subscription business. A renewal model might flag accounts with declining use. The useful response could be better onboarding, a service call or clearer billing information. Automatically sending every flagged customer a discount assumes that price caused the problem. The prediction alone does not establish that.

Why does prediction differ from marketing impact?

Prediction identifies what is likely to happen; causal evaluation asks what changes because the business intervenes.

This distinction matters most when spending money on customers who would act anyway. Targeting highly likely purchasers can produce an attractive response rate while adding little business. Conversely, some customers with moderate purchase likelihood may respond meaningfully to useful information.

Kim and colleagues’ 2023 paper on uplift modelling formalises the difference through treatment effects. The marketing application is to compare outcomes with and without an intervention, rather than treating predicted response as additional response.

There is a second distinction between ranking and calibration. A model may correctly rank stronger prospects above weaker ones while overstating everyone's probability. Scikit-learn’s probability-calibration documentation, checked in October 2026, explains how predicted probabilities can be assessed against observed frequencies.

That matters when a score drives an economic decision. If a team uses estimated probabilities to justify contact costs or incentives, inaccurate probabilities can lead to poor spending decisions even when the ranking looks impressive.

How do you build a useful predictive workflow?

Start with the decision and a simple baseline, protect the evaluation from leakage, then test the action in practice.

1. Define the outcome and decision window

Write what event counts, when it must occur and what information will be available before the decision. Define cancellations, missing records and repeat events. Agree on these rules with the people who will use the score.

For churn, specify whether a customer has explicitly cancelled, failed to renew or simply stopped purchasing within an expected cycle. Those definitions describe different problems. A grocery customer and a furniture buyer should not share the same inactivity rule without a reason.

2. Establish a transparent baseline

Compare the proposed model with a simple rule such as recent activity or purchase frequency. The baseline is not an embarrassment. It establishes what added complexity must improve.

Include maintenance effort in that comparison. A complicated model may require reliable integrations, monitoring and specialist support. A slightly stronger ranking is not automatically worth an operational dependency the team cannot sustain.

3. Prevent information leakage

Use only information that would have existed when the prediction was made. A cancellation reason recorded after cancellation cannot legitimately help predict an earlier cancellation decision.

Scikit-learn’s common-pitfalls guidance explains how leakage can contaminate model evaluation. For time-ordered data, its TimeSeriesSplit documentation highlights why training on future observations and evaluating on the past is inappropriate.

The practical test is to reconstruct the decision as it would have occurred. Keep the final evaluation period separate while developing the model. Repeatedly tuning against the same holdout gradually turns it into training information.

4. Connect the score to a plausible action

Write down why the intervention might help. A reminder can address forgetting. A comparison guide can address uncertainty. A service recovery call can address an unresolved problem. A generic discount may fail to address any of these.

Set contact limits and exclusions. A customer who has asked not to receive marketing should not be brought back into a campaign because a model gives them a promising score. Business rules and permissions remain part of the decision.

5. Test incremental outcomes

Where appropriate, randomly assign eligible customers to the proposed treatment and a suitable comparison. Keep the analysis consistent with that assignment. Compare the outcome that matters after returns, incentives and delivery costs, rather than only opens or clicks.

Choose the observation period before launch and make it long enough for the buying cycle. Also inspect complaints, unsubscribes and service load. A campaign that accelerates purchases without increasing total demand may look different once the full period is observed. The same distinction matters when evaluating AI-driven advertising and automated bidding.

Where does personalisation fit in India?

Personalisation fits best when it removes a real barrier without making unjustified assumptions about the customer.

Language preference, product availability and relevant delivery information can be useful inputs when collected and used appropriately. Treating a postcode as a complete description of income, education or taste is a much weaker inference. Geography can also act as a proxy for factors the business did not intend to use.

Customers outside large metros may face different fulfilment and support conditions. A model trained mainly on metropolitan transactions can confuse those operational differences with lack of demand. Check whether a low conversion score reflects the offer, delivery coverage or how the data was collected.

Explain recommendations in terms the customer can recognise. “Based on the category you selected” is more understandable than an unexplained claim that an offer is uniquely suitable. Avoid inferring sensitive circumstances merely because the data makes it technically possible.

When should a model be refreshed or withdrawn?

A model should be reviewed when the relationship between its inputs, predictions and outcomes changes materially.

A new product, revised pricing, a festival period or a disrupted delivery network may change behaviour. Compare observed outcomes with predictions over time and across relevant segments. A stable overall average can hide a failing segment.

In the first 30 days of a pilot, focus on data completeness, operational exceptions and whether the score arrives in time to support the decision. That is a proposed operating period, not a statistical guarantee. The purchase cycle determines when commercial performance can be judged.

Keep a fallback rule and a way to suspend targeting. If the data feed breaks, the system should not confidently act on stale or empty records.

Who owns predictive marketing?

The marketing decision-owner, data team and customer operations team should share a clearly divided responsibility.

Marketing defines the action and desired outcome. Analysts validate the model and experiment. Data owners maintain inputs and permissions. Customer operations report whether interventions solve problems or create new ones. Finance helps determine whether observed gains remain worthwhile after their costs.

My view

I would begin with a decision the team already makes repeatedly and ask whether better information would change it. That keeps modelling connected to an operational need.

The trade-off is between potential relevance and additional complexity. Personalisation that is difficult to explain, maintain or evaluate can cost more than it contributes. Sometimes the useful result of modelling is discovering that a simpler, clearer offer helps everyone.

A prediction becomes useful marketing only when it leads to an action worth taking.

Start with the counterfactual

Before commissioning a propensity model, write down what would happen without the proposed campaign. Then design the comparison that could distinguish a likely sale from an additional sale. That question will improve the brief for both the analyst and the marketer.