You have 400,000 customers in the database and budget to contact 40,000 of them this month. Sorting by last purchase date is a guess dressed as a method. So is picking the loyalty tier. So is emailing everyone and hoping the unsubscribe rate holds.
A propensity model turns that guess into a ranked list. Every customer gets a number, the list sorts itself, and you draw a line at 40,000.
Here’s the part that gets left out of every article on this topic: the top of that list is usually the least interesting part of it. Your highest-scoring customers were, for the most part, going to buy anyway. Spending your incentive budget on them feels great and moves almost nothing. The deciles worth arguing about sit in the middle, where a well-timed nudge actually changes the outcome. Nobody shows you that, because showing it means showing the output, and almost every page on propensity modeling stops at the definition.
So this one shows the output.
What is a propensity model?
A propensity model is a machine learning model that assigns each customer a probability between 0 and 1 for a specific action inside a specific time window. A score of 0.34 for “purchases within 30 days” means the model estimates a 34% chance that this person buys in the next month.
Both halves of that definition carry weight, and the time window is the half people skip. “Likely to buy” is not a prediction. “Likely to buy within 30 days” is. Change the window to 90 days and you get a completely different model, different features, different scores, and a different marketing decision at the other end. A 12-month window will flag your loyal annual repurchasers. A 14-day window will flag people who are shopping right now. Same customers, opposite lists.
Propensity models are a type of classification model. They sit inside the broader family of predictive analytics methods that use historical patterns to estimate what happens next.
The four propensity models marketing teams actually use
Propensity modelling (the British spelling shows up constantly in UK and Australian search results, same concept) covers a family of models rather than one technique. Four of them account for most real marketing work.
Propensity to buy. Target variable: did this customer make a purchase in the window? Used by growth, ecommerce, and CRM teams. The score feeds campaign targeting, contact budget allocation, and offer selection.
Propensity to churn. Target: did this customer stop buying or cancel inside the window? Used by retention and customer success. The score triggers intervention while there’s still time to act, which is the entire point of churn propensity modeling over churn reporting.
Propensity to convert. Target: did this lead or trial user become a paying customer? Used by demand gen and RevOps for lead scoring and sales prioritization. The score decides who gets a human and who gets an email.
Propensity to upgrade. Target: did this customer move to a higher tier or add a product? Used by CS and account teams. This one overlaps with cross-sell and upsell propensity, which deserves its own treatment.
Same math underneath all four. Completely different business decisions on top.
The propensity to buy model, step by step
Walk one end to end. An outdoor gear retailer wants to know who will buy in the next 30 days.
The outcome. One purchase of any product, any channel, within 30 days of the scoring date. Written down, agreed on, unambiguous.
The observation period. Twelve months of behavior leading up to each scoring date. Long enough to capture seasonality, short enough that ancient behavior doesn’t dominate.
The training setup. Pick a date in the past, say March 1. Look backward twelve months to build features. Look forward 30 days to see who actually bought. That forward window is the label. Repeat across several historical scoring dates so the model sees more than one month’s worth of conditions.
The features. Days since last order. Order count in the last 90 days. Average basket. Category mix. Email engagement trend. Site sessions in the last two weeks. Return rate. Whether the last order used a discount. Thirty to a few hundred of these, most of them derived from data the retailer already has.
The output. Every active customer gets a number. 0.61 for the person who has browsed jackets four times this week. 0.02 for the person who bought once in 2023 and never opened an email since.
That’s the model. What happens next is where the money is.
How a propensity model works under the hood?
No code here, just the shape of it.
The model looks at thousands of past customers who did and didn’t take the action, and learns which combinations of behavior separated the two groups. Not single variables. Combinations. Recent site visits alone mean little. Recent site visits from someone who bought twice last quarter and has never returned an item means quite a lot.
That learning happens through classification models, most commonly gradient boosted trees for tabular customer data, sometimes logistic regression when interpretability matters more than accuracy.
Validation is where propensity models earn or lose your trust. The model gets trained on one slice of history and tested on a slice it never saw, ideally a later time period rather than a random sample. If it scores 0.85 AUC on training data and 0.61 on the holdout, it memorized rather than learned. Pecan runs this validation automatically and checks for the failure modes that quietly wreck propensity models, including data leakage and unbalanced labels.
Then the live base gets scored, and the scores refresh on whatever cadence your campaigns run on.
What data you need (and what you can skip)
Most teams already have everything on this list sitting in their warehouse.
| Data type | Examples | What the model uses it for |
| Transactions | Order date, amount, items, channel, discount used | Recency, frequency, spend patterns, category affinity |
| Behavior | Sessions, page views, cart adds, email opens and clicks, app opens | Intent signals in the weeks before a purchase |
| Customer attributes | Signup date, acquisition source, tier, region | Tenure effects and segment differences |
| Service history | Tickets, returns, refunds, chat contacts | Friction that suppresses or delays the next purchase |
What you can skip: purchased demographic appends, which rarely add signal over behavior you already observe. Survey data with low coverage. Anything recorded after the outcome you’re predicting, which is the fastest way to build a model that looks brilliant and predicts nothing.
Data that’s inconsistent, partially missing, or spread across systems is the normal starting condition, and it’s workable. Pecan connects to Snowflake, BigQuery, Redshift, and Databricks and handles the joins, time windows, and feature preparation as part of the build.
Reading the scores: what to do with each decile
Here’s the part that almost never makes it into an article. You have a scored list. Now what?
Split the base into deciles by score. The shape below is typical for a 30-day purchase model on a consumer base with a 4% baseline rate. Treat the numbers as illustrative, since yours will differ, though the pattern holds remarkably well across companies.
| Decile | Score range | Share of all predicted buyers | What to actually do |
| 1 (top 10%) | 0.42 to 0.91 | ~35% | Contact, no incentive. They’re already coming. A discount here is margin you’re handing away. |
| 2 | 0.24 to 0.42 | ~20% | Standard promotion. Make the path short. |
| 3-4 | 0.11 to 0.24 | ~22% | Your real target. Test incentives against a holdout. This is where a nudge changes outcomes. |
| 5-6 | 0.05 to 0.11 | ~13% | Test a stronger offer on a small cell before scaling. |
| 7-8 | 0.02 to 0.05 | ~7% | Low-cost channels only. No paid retargeting spend. |
| 9-10 | below 0.02 | ~3% | Suppress. Every contact here is budget that decile 3 could have used. |
Read that table again with your incentive budget in mind. The top decile carries the most predicted buyers, which is exactly why discounting them is expensive: you’re paying people to do what they’d have done for free. Deciles three through six carry a third of your predicted conversions and contain the customers whose decision is genuinely undecided.
Standard practice is to hammer the top and ignore everyone else. That produces excellent-looking campaign reports and mediocre incremental revenue.
Propensity model vs propensity score matching
Quick disambiguation, because these two share a word and nothing else.
A propensity score model in causal inference is a statistical technique for estimating treatment effects from observational data. Researchers use it to match treated and untreated subjects with similar characteristics, approximating a randomized trial when they can’t run one.
A propensity model in marketing predicts whether an individual will take an action. Different purpose, different output, different field. If you searched for one and landed on the other, that’s the difference in three sentences.
Building your own vs using a platform
Two honest paths, and the right one depends on what you already have.
Building in house means a data scientist or two, a feature pipeline, a training and scoring schedule, monitoring for drift, and someone who owns the whole thing when it degrades. For a company with an established data science function and a long-term modeling roadmap, this is entirely reasonable. The models are well understood and the tooling is mature.
The cost people underestimate is not the first model. It’s model eleven, and the retrain schedule, and the fact that the person who built model three has left.
Using a platform means connecting your data and getting scores. This is what we do. With Pecan, you describe the business question in plain language, the Predictive AI Agent handles data preparation, feature engineering, model selection, and validation, and predictions land in Salesforce, HubSpot, or your warehouse where campaigns can act on them. No coding required, and no data science team needed to keep it running.
We’re not going to claim this is the right answer for everyone. If your team already runs production ML and has capacity, build it. If your marketing team is waiting six weeks in a data science queue for a scored list, that queue is the actual problem, and a platform solves it faster than headcount will. Book a demo and we’ll walk through how it would work on your use case. Teams focused specifically on retention can also start with our customer churn solution.
Where propensity models go wrong
Four failure modes account for most disappointing propensity projects.
Target leakage. A feature that only exists because the outcome happened. Shipping address updated, loyalty points redeemed, order confirmation email opened. The model finds these instantly and scores beautifully in testing, then collapses in production. Any feature that correlates suspiciously well deserves a hard look at when it gets recorded.
A window that doesn’t match your campaign cycle. A 90-day model driving a weekly email program will feel permanently wrong, because it is. The prediction window should match the period over which your action can plausibly influence behavior.
Scoring the base and changing nothing. The most common outcome by a wide margin. A model gets built, a dashboard gets made, and the campaign calendar runs exactly as it did before. Prediction without a change in what anyone does is an expensive report.
Letting the model go stale. Customer behavior shifts. Pricing changes, a competitor launches, a season turns. A model trained on last year’s patterns degrades quietly, without an error message. Retraining cadence should be decided at build time, not discovered after a bad quarter.
What is the meaning of propensity model?
What is an example of a propensity model?
Is logistic regression a propensity model?
What is the difference between a propensity model and a predictive model?
Dror is the VP of Data and Analytics at Pecan AI, where he leads the analytics strategy that powers both customer success and Pecan’s own growth. He joined Pecan as Director of Analytics after years of data leadership roles across tech and fintech, bringing a firsthand understanding of what it takes to make data actually useful for business teams.
