Challenge · 01
Long-tail SKUs had so little signal the legacy model defaulted to safety stock for years.
Resolution
Hierarchical shrinkage — the long tail borrows strength from its category × region group instead of being rounded up.
← All case files
Case / 2025.04 · Retail
Aisle 19
Hierarchical demand forecasting, long-tail SKUs
A forecast that respects the shape of the catalog — fast-movers learn from themselves, the long tail borrows strength from its category and region.
/Retail-Group/ · Shipped
−18% inventory holding · forecast variance halved on long-tail SKUs.
/ 01 · Stakes
What was on the line.
A 1,200-store multi-format retailer was carrying about 25% more stock than it needed because the legacy forecast couldn’t tell pasta from imported truffle salt. Top sellers were predicted well; everything in the long tail got rounded up to safety stock.
The CFO saw the inventory line; merchandising saw stockouts on the same SKUs the warehouse was overstocked on. Category managers had stopped trusting the system and were overriding orders by gut feel.
/ 02 · Approach
How we built it, in five moves.
- 01
Hierarchical fit at SKU × store × season
The forecast learns at three grains at once. Sparse long-tail SKUs shrink toward their category × region group, with shrinkage strength learned per group.
- 02
Causal signal layer with lag alignment
Weather, macro, promo calendar and regional events feed in with lags searched per category — a hot weekend lifts beer the same day and ice cream three days later.
- 03
Reconciliation for coherence
A MinT-style reconciliation step makes store totals add up to chain totals, so one number rolls up cleanly from a single jar of olives to the Q3 plan.
- 04
Per-SKU drift, targeted retraining
A per-SKU drift score flags only the SKUs that changed; they retrain in hours without touching hierarchy weights.
- 05
Constraints at the recommendation seam
Order cutoffs, MOQs, case packs and shelf life live in a separate recommendation layer, so a new supplier rule never poisons the model fit.
/ 03 · Challenges
What got in the way, and how we resolved it.
Challenge · 02
Category managers didn’t trust the model and routinely overrode it.
Resolution
A what-if dashboard exposing each signal’s contribution per SKU. Trust came from visibility; adoption hit 90% inside six months.
Challenge · 03
Drift was constant, and full retrains lagged it by weeks.
Resolution
Online per-SKU drift detection triggers targeted retraining; full-hierarchy refits stay on a quiet monthly cadence.
/ 04 · Outcomes
What changed, in numbers.
- −18%
- Inventory holding
- 50%
- Long-tail variance cut
- 90%
- Category-manager adoption
- 1,200
- Stores live
- Inventory holding down 18% across the catalog within two quarters
- Forecast variance on long-tail SKUs halved against the legacy baseline
- Category-manager adoption at 90% inside six months, mostly via the what-if dashboard
- Full rollout across 1,200 stores on one live banner, with two more in pilot
/ 05 · Evidence
Live across 1,200 stores on one banner with two more in pilot. Per-SKU drift retraining runs continuously; full-hierarchy refit monthly.
- Hierarchical forecasting
- MinT reconciliation
- Causal signal layer
- Per-SKU drift detection
- What-if dashboards
Next case file · Platform · DevOpsNeed to Know — ABAC-gated RAG for an internal knowledge fabricOpen file
