Where RFM Score sorts customers into descriptive tiers based on past behavior, purchase propensity score does something structurally different: it predicts forward, estimating the probability a specific customer buys again within a stated window (commonly 30, 60, or 90 days). That distinction matters operationally — RFM tells you who's been valuable, propensity tells you who's about to act.
Full implementations use logistic regression or an equivalent classification model trained on historical purchase data, but a defensible propensity score doesn't require machine learning to start. A weighted-sum model over the same behavioral signals RFM already uses produces a workable first version, which is the practical entry point for most marketing teams before investing in a full predictive model.
The score only means something relative to a stated time window and a validated threshold — a 0.7 propensity score is meaningless without knowing what window it predicts and whether that cutoff has actually been checked against real purchase outcomes, the same discipline RFM segments require but that propensity scores get away with skipping more often because the single number feels more authoritative than it is.
Formula
PPS = (w1 × Recency_norm) + (w2 × Frequency_norm) + (w3 × Engagement_norm) + (w4 × Monetary_norm)
Each input normalized 0-1, weights summing to 1. For retention-focused use cases, recency and frequency typically carry the highest weights. This weighted-sum version is a practical starting point; a full implementation replaces it with a logistic regression trained on historical conversion outcomes, which produces a genuine probability rather than an ad hoc weighted index.
Why Purchase Propensity Score matters
A propensity score lets you route budget and messaging by predicted near-term action instead of past value alone — sending a high-propensity customer a low-friction nudge instead of a discount they didn't need, and reserving heavier incentives for customers whose score says they're on the fence.
Where propensity and RFM disagree
A customer with a high RFM score — frequent, recent, high-spend — can carry a low purchase propensity if their last few sessions show declining engagement and no cart activity, a genuine early churn signal RFM's backward-looking tiers won't catch. Conversely, a mid-tier RFM customer showing a sudden spike in browsing and email engagement can carry a high propensity score despite unremarkable purchase history — exactly the customer a retention team would otherwise miss by segmenting on RFM alone.
Common mistakes
Treating the score as accurate without validating the threshold
A 0.7 cutoff chosen because it 'feels high' is unvalidated. Check it against actual outcomes — of customers scored above your cutoff last period, what share genuinely purchased in the stated window — before using the threshold to drive spend decisions.
Leaving the time window unstated
A propensity score without a defined window (30 days? 90 days?) can't be compared across models or audiences, and the same customer's 'true' propensity looks completely different at 30 versus 90 days.
Using it as a replacement for RFM rather than alongside it
The two answer different questions — past value versus predicted near-term action — and the customers where they disagree are frequently the most operationally interesting segment, not a discrepancy to resolve by picking one metric.
Where we work on this