Loading data…
Monthly revenue
GBP, purchases only (returns excluded)What this page shows: how revenue is spread across customers at the 9 Sep 2011 cutoff, and how often each group bought again in the next 91 days.
RFM segments
Segment table
Revenue share is of revenue up to the cutoff| Segment | Customers | Revenue share | Bought again | Median days since last order | Median orders |
|---|
How to read this: "Bought again" is the share of the segment with at least one order from 10 Sep to 9 Dec 2011. It is observed, not predicted.
k-means clusters
| Cluster | Customers | Share of customers | Revenue share | Median days since last order | Median orders | Median revenue | Bought again |
|---|
How to read this: the clusters come from log recency, frequency and spend, standardised. They give a second, data-driven grouping to set beside the rule-based RFM segments.
Revenue concentration
Cohort retention
Share of each first-purchase cohort buying in month kWhat this page shows: how well the churn model and the CLV model did on the 91 days after the cutoff, which neither model saw when it was fitted.
Churn model metrics
Out-of-time test unless the setting says otherwise| Model | Setting | Customers | Churn rate | ROC AUC (95% CI) | PR AUC | Brier |
|---|
ROC curve, test window
Logistic model vs the recency ruleHow to read this: each point is a cut-off. Higher and further left is better. The diagonal is a coin flip.
Churn by risk decile
Decile 1 = highest predicted riskCalibration
Mean predicted vs observed churn, by decileTop features
Drop in test AUC when the feature is shuffledHow to read this: a longer bar means the model leans on that feature more. Gaps between purchases and time since the last one matter most.
CLV check: revenue by predicted decile
CLV model vs simple baseline
91-day holdout, all customers seen by the cutoff| Model | Target | Total error | MAE per customer | Pearson r | Spearman r |
|---|
How to read this: the baseline repeats each customer's past rate. BG/NBD + Gamma-Gamma ranks customers better (higher correlation) but undershoots the total, partly because the holdout is the busy autumn season.
What this page shows: the expected profit of sending a voucher to the top customers under each ranking. Move the sliders to change the assumptions; the curves are recomputed for all 4,305 scored customers in your browser.
Assumptions
assumed5%25%
0%20%
£0.25£5
20%40%
25%100%
nobodyeveryone
Save rate: share of would-be churners the offer keeps. Customers who would have bought anyway also use the voucher, so its cost falls on them too.
Expected profit by targeting depth
GBP, top x% of customers by each ranking; dashed line = your depthHow to read this: each line adds customers in its ranking order. The curve rises while the next customer's expected gain beats the voucher and contact cost, then falls. Ranking by CLV alone picks loyal customers who would have bought anyway, so every voucher is a loss.
Rankings compared
Current slider settings| Ranking | Best depth | Best profit | At your depth |
|---|
A/B test size
Two-sided test, 5% significance, 80% power| Save rate to detect | Control churn | Offer churn | Per arm | Total |
|---|
What is assumed and what is observed
| Input | Type | Where it comes from |
|---|