CASE 13
12 weeks
Fewer stockouts, less overstock, fewer spreadsheets.
A multi-category retailer forecasted demand by pasting last year's sales beside a gut-feel column. We built a SKU-level forecasting pipeline with hierarchical reconciliation, delivered suggestions into the system where purchase orders actually get written, and published accuracy back-tests so the category team could see exactly when to trust it.
THE PROBLEM
A category manager forecasted demand weekly by pasting last year's sales next to 'gut feel' columns. Peaks meant stockouts; lulls meant writedowns.
The manual process worked acceptably for top sellers, where the manager's intuition was genuinely good, and badly for the long tail, which was too large to think about SKU by SKU.
Promotions and seasonality were handled by memory. If the person who ran last year's campaign had left, the demand spike was a surprise for the second time.
Category totals and SKU-level plans never agreed, so finance planned against one number and the warehouse ordered against another.
CONSTRAINTS WE WORKED UNDER
WHAT WE DID
Built a SKU-level forecasting pipeline (hierarchical reconciliation so category totals always match).
Mixed a statistical baseline with a trained neural model for the long-tail SKUs.
Integrated with the warehouse management system so reorder suggestions land where purchase orders actually get written.
Weekly back-tests and accuracy dashboards so the category team sees exactly when the model helped and when it didn't.
ARCHITECTURE
01
dbt models over sales, promotions, price changes, stockout periods and calendar effects. Critically, stockouts are masked rather than read as zero demand — otherwise the model learns to under-forecast exactly the SKUs that ran out.
02
A statistical baseline handles seasonal, well-behaved SKUs; a trained neural model handles the intermittent long tail where classical methods degrade. Selection is per SKU, decided by back-tested accuracy rather than by preference.
03
SKU forecasts reconciled upward so subcategory and category totals sum correctly. This removed the recurring argument between finance and the warehouse about which number was real.
04
Reorder suggestions written into the warehouse system with the forecast, the confidence interval and the reasoning attached. The buyer accepts, edits or rejects in the tool they already use, and every decision is captured.
05
Weekly rolling-origin back-tests reporting accuracy by category, by SKU tier and against the previous manual process. Published whether the numbers flatter the model or not.
KEY DECISIONS
WHY
Zero sales during a stockout is censored data, not low demand. Ignoring this is the most common way retail forecasting quietly fails.
TRADE-OFF
Less usable history for the SKUs that stocked out most, which is exactly where more data would have helped. Correct anyway — biased data is worse than less data.
WHY
Two years of anomalous behaviour would have taught the model patterns that will not recur. Attempting to model it would have burned weeks for negative value.
TRADE-OFF
A gap in history and weaker long-horizon seasonality for some categories. We documented the choice so nobody rediscovers it as a bug.
WHY
Adoption is a delivery problem more than a modelling one. A better forecast that lives somewhere buyers do not go improves nothing.
TRADE-OFF
WMS integration was fiddlier than a standalone app and added roughly two weeks. It is the reason the model changed actual orders.
THE HARD PARTS
Intermittent-demand SKUs are genuinely hard, and the honest answer was that point forecasts are the wrong tool. We shifted those to service-level-driven reorder points with explicit uncertainty, which is less impressive and considerably more useful.
Historical promotions lived partly in the ERP and partly in a marketing spreadsheet with different SKU identifiers. Reconciling them took longer than training the models and mattered more to accuracy.
The first back-test showed the model losing to the manager on top-tier SKUs. We published that rather than hiding it, and scoped the rollout to the long tail where the win was large. Top-tier came later, on evidence.
HOW IT RAN
01
dbt models over sales and inventory history, stockout and promotion reconstruction, and an honest assessment of which SKUs had enough signal to forecast at all.
02
Statistical baseline plus rolling-origin back-tests benchmarked against the manual process — the comparison that mattered.
03
Trained model for the long tail, per-SKU model selection, and hierarchical reconciliation across the category tree.
04
Reorder suggestions into the WMS with confidence intervals, accuracy dashboard, and a phased rollout starting with the long tail.
OUTCOME
Down
Stockouts on A-tier SKUs, meaningfully
Double-digit %
Reduction in inventory holding cost
Weekly
Published back-tests, including the losses
In-WMS
Suggestions delivered where POs are written
Stockouts down meaningfully on A-tier SKUs. Inventory holding cost down a double-digit percentage, freeing cash.
Freed working capital was the number the CFO cared about, and it came mostly from the long tail — thousands of small overstocks nobody had the time to think about individually.
The category manager's role shifted from producing forecasts to overriding them where their knowledge genuinely beat the model. The back-test dashboard tells them where that is.
STACK
DATA
MODELLING
SERVING
DELIVERY
HANDOVER
