Personalisation & ML · ML
Grow repeat basket size with purchase-sequence models
+22%
Hand-picked product pairings encode what the merchandiser believes, not what customers actually buy next. This engagement learns the purchase sequence from order history and lets the model choose the follow-on offer.
The process
Rebuild the order sequence
Order lines reshaped in the warehouse into per-customer sequences, with returns, cancellations and replenishment cycles handled explicitly rather than ignored.
Separate repeat from discovery
Consumable SKUs that simply recur split from genuine cross-category moves, so the model is not rewarded for predicting the obvious.
Train and validate the next-basket model
A sequence model trained on earlier periods and validated on held-out later ones, benchmarked against the current hand-picked rules as the baseline to beat.
Ship it into one channel first
Predictions delivered to a single surface — post-purchase email or the account page — with a margin and stock filter before anything is shown.
Measure basket, not clicks
Repeat basket size, order frequency and margin read against a holdout group, with returns netted out before the result is claimed.
What we need from you
- Twelve or more months of transaction history at line level
- Product catalogue with category and margin
- One channel able to render dynamic recommendations
What you get
- Customer purchase-sequence dataset
- Validated next-basket model with baseline comparison
- Live recommendations in one channel, read against a holdout
Timeline
Six to ten weeks, data quality permitting.