Retail Analytics · Retention Intelligence

Turn every order into a retention, service, and expansion decision.

Marketplace Compass reads raw transactions and returns the three answers operators act on: who is about to churn, which segment to serve next, and where the next dollar of growth lives — each traceable to the data and honest about model limits.

How the models workDec 2009 – Dec 2011 · UCI Online Retail II (CC BY 4.0)

Customers

5,878

unique buyers

Orders

36,969

invoices

Revenue

$17.7M

gross, all-time

Repeat rate

72.4%

≥2 orders

Model ROC-AUC

0.783

baseline 0.788

Model card

Repeat-purchase model

Predicts whether a customer buys again within 90 days of the cutoff. Evaluated on a strict temporal holdout (no shuffle, no leakage).

Temporal holdout · n=5,249

Metric	HGB (model)	Logistic (baseline)
ROC-AUC	0.783	0.788
PR-AUC	0.751	0.744
Brier (lower is better)	0.195	0.193
Top-decile lift	2.14×	2.09×

Honest read: the logistic baseline slightly edges the gradient-boosted model on ROC-AUC (0.788 vs 0.783, a 0.005 gap), while HGB wins on PR-AUC (0.751 vs 0.744) — the metric that matters most for ranking who to retain. Both models beat a 42.7% base rate and lift the top decile ~2.14×. On this dataset a well-regularized linear model is a genuinely strong baseline; we report it rather than hide it.

What drives the prediction

Permutation importance on the test set. Recency dominates; frequency and monetary follow — a textbook RFM signal, measured not assumed.

Permutation importance = mean drop in test ROC-AUC when a feature is shuffled. Near-zero or negative values (e.g. Is UK, Unique products) carry little signal.

Customer segments (RFM)

5,878 customers scored on recency, frequency, and monetary value, then labeled into five actionable segments as of December 9, 2011.

Segment	Customers	% base	Avg spend	Avg orders	Pred. repeat
ChampionsRecent, frequent, high-spend — reward and ask for referrals.	661	11.2%	$14,568	24.4	87.9%
LoyalSteady repeat buyers — upsell and protect the relationship.	1,184	20.1%	$3,457	8.1	62.7%
At-RiskFormerly valuable, now lapsing — win back before they churn.	688	11.7%	$3,359	7.2	22.6%
HibernatingLong dormant, low value — low-cost reactivation only.	2,237	38.1%	$468	1.7	10.4%
PotentialRecent but light — nurture toward a second and third order.	1,108	18.8%	$599	2.2	39.7%

Segment size vs. value

Many customers sit in low-value segments; a small Champions core carries outsized spend.

Bar height = customers in segment; label above each bar = average lifetime spend per customer.

What is computed vs. proposed

Computed here (measured results)

Customer-level repeat-purchase / churn prediction (this pipeline, temporal holdout).
RFM segmentation (Champions / Loyal / Potential / At-Risk / Hibernating).
90-day CLV proxy tiers (Platinum/Gold/Silver/Bronze).

Proposed (needs Olist data)

Order-level late-delivery risk model (needs Olist delivery timestamps).
Brazil geospatial expansion / regional demand model (needs Olist geolocation).

This demo runs on UCI Online Retail II (CC BY 4.0) — a keyless, directly downloadable dataset. Olist Brazilian E-Commerce is the richer PRIMARY target: its delivery timestamps and geolocation would unlock order-level late-delivery risk and a Brazil geospatial expansion model — but it is Kaggle-gated (CC BY-NC-SA 4.0) and not used here. The substitution is stated plainly rather than papered over.