Executive summary
Growth
Monthly GMV grew from R$137K in Jan-2017 to a Black-Friday peak of R$1.18M in Nov-2017 and plateaued around R$1.0-1.2M per month through 2018 (Jan-Jul 2018 vs Jan-Jul 2017: +160%). Health Beauty is the largest category at 8.8%; the top 10 categories hold 62% of GMV. Credit cards carry 78% of payment value and the average order is split into 2.9 installments.
Customers: retention, segments, lifetime value
Retention is the structural weakness. Average month-1 retention is 0.48% and month-6 is 0.26%; 32% of 'second orders' happen the same day (split checkouts, not loyalty) and the median gap to a true second order is 25 days. RFM segmentation shows 12,971 lapsed high-value customers holding 28% of historical GMV - the natural win-back list. A BG/NBD model (holdout error +12% on total repeat purchases) expects only 1,824 repeat purchases from the whole base in the next 12 months; the top CLV decile holds 38% of predicted value.
Delivery performance and customer experience
Median delivery is 10.6 days against a 25-day promise, so most orders arrive early - and still 6.7% miss the date, peaking at 19% in 2018-03 (carrier capacity, then the May-2018 truckers' strike). Late orders lose time in transit (median 27 vs 7 days), not at the seller. Controlling for order value, distance and basket size, lateness dominates every other driver of bad reviews; multi-seller baskets are the second-largest driver (3.5x odds).
Seller economics
2,878 sellers sold in the window; the median seller made R$979 and 19% sold exactly once. A composite health score (50% volume, 25% on-time rate, 25% reviews) tiers sellers A-D; 19 sellers with 50+ orders combine a >=12% late rate with reviews <=3.8 and account for R$279K of GMV - the enforcement list. Seller churn mirrors customer churn: 57% of a cohort sells again the next month and 44% six months later.
Seller acquisition funnel (B2B marketing)
8,000 marketing-qualified leads became 842 closed deals (10.5%) in a median 14 days, and 380 of those sellers (45%) made at least one sale, a median 44 days after signing. Measured on GMV per lead rather than close rate, paid search (R$115) and organic search (R$103) outperform social (R$38, close rate 5.6%) by ~3x. Value is concentrated: the top 5 activated sellers produced 32% of all funnel GMV. 14% of leads have no recorded source - a tracking gap worth closing before re-allocating budget.
Geography
Sao Paulo state is 37% of GMV and 71% of orders ship from SP sellers; 65% of orders cross a state line. The North region is 2.6% of GMV but waits 24 days on average and pays freight worth 51% of the product price (SP: 9 days, 25%). The Northeast has the highest late rate at 14%.
Weekly GMV forecast
Three models were compared on a 6-fold rolling-origin backtest with an 8-week horizon. The 'naive_last4w_mean' baseline won (mean WAPE 14.3%); Holt-Winters damped trend was close (17.0%) and log-SARIMA over-forecast (+24% bias) because it extrapolated 2017 growth into the 2018 plateau. Honest takeaway: with 85 weeks of history and a regime change, a well-chosen naive baseline is hard to beat; the value of the exercise is the backtest harness, not the point forecast. Next 8 weeks: ~R$2.44M.
Predictive models
Late-delivery risk (operational): gradient boosting on checkout-time features with point-in-time seller history, trained to Feb-2018 and tested forward on Mar-Jul 2018: AUC 0.74, PR-AUC 0.18 vs a 7.2% base rate. The riskiest decile is late 20% of the time (2.8x) and the top two deciles capture 49% of all late orders - enough to drive proactive customer notification. Repeat-purchase propensity (strategic): deliberately reported as a weak result - AUC 0.57. First-order attributes barely predict who comes back, which says the retention problem is structural (assortment, one-off purchases) rather than a targeting problem.
Recommendations and how to test them
- Proactive delay management. Score every order at checkout with the risk model; for the top 20% risk, pad the promised date and send a proactive ETA update with a small voucher when the carrier scan slips. Target: cut 1-2 star reviews among flagged orders by 25%.
- Seller SLA and tiering. Publish the A-D health score to sellers; require <=2-day handling for C/D tiers and route the 19 problem sellers to a remediation plan or delisting.
- Re-allocate B2B acquisition spend from social to search on a GMV-per-lead basis, and fix source tracking on the 14% of leads with unknown origin first.
- Win-back the 12,971 lapsed high-value customers with a holdout-measured voucher campaign; stop paying to re-acquire the Hibernating segment (24% of customers, 11% of GMV).
- Reduce multi-seller baskets' failure rate (second-largest review driver) by consolidating shipments or surfacing per-seller ETAs at checkout.
Each recommendation maps to a pre-registered experiment. Sample sizes below use the 2018 run-rate of 1,456 orders/week.
| Test | Primary metric | Baseline | MDE (relative) | n per arm (80% power, α=0.05) | Weeks at 2018 volume | Feasible in 12 weeks? |
|---|---|---|---|---|---|---|
| T1 Proactive delay notification Telling customers whose order is flagged high-risk (top 2 risk deciles) that it may arrive late, with a new ETA and a small voucher, cuts 1-2 star reviews among flagged orders. | detractor rate among flagged orders | 18.6% | -25% | 988 | 7 | yes |
| T2 Seller shipping SLA (handling <= 2 days) Enforcing a 2-day handling SLA for B/C/D-tier sellers reduces late deliveries. | late delivery rate | 7.2% | -20% | 4,586 | 11 | yes |
| T3 Win-back campaign for Lapsed High-Value A 10% voucher emailed to Lapsed High-Value customers lifts 60-day repurchase vs a holdout. | 60-day repurchase rate | 1.2% | +50% | 6,379 | yes | |
| T4 Free-shipping threshold (R$150) Free shipping above R$150 raises AOV without eroding contribution. | average order value | R$166 | +5% | 11,926 | 17 | no - needs longer run or bigger MDE |