What is propensity modelling and how do marketing teams use it?
Most marketing teams are spending money they cannot justify on audiences they cannot define. They run broad campaigns, apply rough segmentation and then attribute whatever comes back to the last click that fired. The model is familiar. It is also quietly expensive.
The problem is not effort. Most marketing directors running ecommerce or technology-led businesses are not short of data or ambition. The problem is that they are making probabilistic decisions - who to target, when to contact them, what to offer - using tools that were designed to describe the past rather than anticipate the future.
Propensity modelling is the discipline that changes that equation. It replaces intuition and segment-level generalisations with individual-level predictions: the likelihood that a specific customer will buy, churn, upgrade or respond. Used properly, it shifts marketing from broadcast to precision.
This article explains what propensity modelling actually is, how it works in practice, where marketing teams commonly go wrong with it and how to use it to drive measurable commercial outcomes.
What propensity modelling actually means
A propensity model is a statistical model that estimates the probability of a specific behaviour occurring. It outputs a score - typically between zero and one - for each individual in a population. The higher the score, the more likely that person is to take the action you are predicting.
The behaviour could be almost anything: purchasing within the next 30 days, cancelling a subscription, responding to an email, upgrading from a free to a paid tier, lapsing after a promotional period. What matters is that the behaviour is clearly defined, historically observable and commercially relevant.
The inputs are usually a combination of behavioural, transactional and demographic signals. For an ecommerce business this might include purchase frequency, recency, basket value, product category affinity, browsing behaviour, email engagement and on-site search history. For a SaaS business it might include login frequency, feature adoption depth, support ticket volume and billing history.
The model learns which combinations of signals predict the target behaviour in your specific customer population. A customer who browsed a product category three times in seven days, opened two emails and last purchased 45 days ago may carry a 0.78 propensity to purchase in the next fortnight. Another customer with superficially similar demographics but different behavioural signals may score 0.12.
That distinction - made at scale, automatically - is what separates propensity-led marketing from everything else.
The three decisions propensity scores should drive
Marketing teams that use propensity modelling well do not treat scores as an analytics curiosity. They wire them into three categories of commercial decision.
Who to target. The most direct application. Rather than mailing your entire database or applying a blunt recency-frequency-monetary (RFM) segment, you rank your population by propensity score and focus spend and effort on the top deciles. A consumer subscription business might find that their top two propensity deciles account for 60% of all conversion events despite representing 20% of their addressable audience. Targeting them exclusively, or weighting them heavily in paid media, materially improves return on ad spend without requiring a larger budget.
What to offer. Propensity scores can be built for multiple outcomes simultaneously. A retailer running a promotional campaign should know not just who is likely to buy, but who would have bought anyway. Offering a discount to a customer with a 0.85 baseline purchase propensity does not drive incremental revenue - it destroys margin. Propensity modelling lets you separate the persuadables from the certain-to-buys and the unlikely-to-buys, and tailor the offer accordingly.
When to contact. Timing models estimate the probability of conversion across different time windows. A B2C subscription business with a 30-day free trial knows that the propensity to convert shifts dramatically across the trial lifecycle. A customer who activated two features in the first three days has a different score trajectory from one who logged in once on day one and not since. Propensity-informed timing means the retention email lands when it will actually matter - not on day 28 because the sequence said so.
Where most marketing teams go wrong
The most common failure is not a modelling failure. It is a deployment failure.
Teams invest in building a model, validate it satisfactorily on historical data and then fail to connect it to anything that changes a decision. The scores sit in a data warehouse. The campaign team continues to use the same segments they used before. The model becomes a PowerPoint slide.
A second failure mode is modelling the wrong behaviour. A large fashion retailer we have worked alongside built a purchase propensity model that performed well on accuracy metrics but predicted any purchase across their entire catalogue. The output was not actionable because the business needed to know which customers were likely to buy in a specific high-margin category, not whether they would purchase at all. The model answered a question no one actually had.
A third failure is neglecting model decay. A propensity model trained on pre-pandemic behavioural data and still running in production three years later is not just stale - it is actively misleading. Customer behaviour shifts. New product lines change purchase patterns. A model that is not monitored and retrained on a regular cycle will quietly degrade without anyone noticing until the campaign results start to slide.
Avoiding these failures requires three things: a clearly defined commercial question before any modelling begins, a direct integration pathway from score to activation channel and a production monitoring regime that flags when model performance drops below an agreed threshold.
Building a propensity model: what a practical process looks like
You do not need a team of twenty data scientists. A well-scoped propensity model can be built, validated and deployed in four to six weeks by a small technical team with access to clean behavioural data. The steps are consistent across most marketing applications.
- Define the target event. Specify exactly what behaviour you are predicting, over what time horizon and in which customer population. "Purchase within 21 days among customers who have not bought in 90 days" is actionable. "Likelihood to engage" is not.
- Audit your training data. Identify what signals are available, how far back they run and whether they are complete enough to be useful. Missing data in the most predictive features is a common problem and needs to be surfaced early.
- Build and validate the model. Logistic regression remains competitive for many marketing propensity problems. Gradient boosted models typically outperform on complex behavioural data. Validate on a holdout set; assess calibration as well as discrimination - a model that ranks customers correctly but systematically overstates probabilities will mislead your campaign planning.
- Define score-to-action logic. Before deployment, specify exactly what changes in each channel when a customer moves from one score band to another. This is the step most teams skip. Without it, the model has no commercial impact.
- Monitor and retrain. Set a cadence - quarterly at minimum - for evaluating whether the model's predictions still align with observed outcomes. Track score distribution drift as well as predictive accuracy.
Tools like Rodan's Quantsole platform can surface propensity-derived signals directly into commercial reporting, allowing marketing and trading teams to query scores, track cohort behaviour and act on outputs without requiring a data team intermediary for every decision.
From scores to commercial outcomes
Propensity modelling is not a technical project. It is a commercial one that requires technical execution. The return only materialises when the scores reach the people and systems that make targeting, timing and offer decisions - and when those decisions change as a result.
A growth-stage ecommerce business running a loyalty programme should not be offering the same re-engagement incentive to every lapsed customer. Some will return with a light-touch email. Some need a meaningful offer. Some have already moved to a competitor and are unlikely to return regardless of what you spend. Propensity scoring separates these groups cleanly and lets you allocate budget accordingly.
The cost of not doing this is not abstract. It is measurable: in wasted media spend on audiences who were already converting, in margin eroded by discounts offered to customers who did not need them and in lifetime value left on the table because the timing of the right message was wrong by two weeks.
If you are running marketing at scale without individual-level propensity scores feeding into your targeting and personalisation stack, you are competing at a structural disadvantage. The businesses pulling away from the field are not doing so because they have bigger budgets. They are doing so because they are making better predictions about individual customer behaviour - and acting on them faster.
If you want to understand what your customer data can actually support, a Rodan diagnostic engagement is the right starting point. We will assess your data assets, identify the highest-value modelling opportunities and give you a clear roadmap for putting propensity to work. Book a diagnostic at rodan.io.
