Elicon
← ALL REFERENCE ENGAGEMENTS
REFERENCE ENGAGEMENT
Retail

Demand Forecasting for a Multi-Category Retailer

SKU-level forecasts delivered where purchase orders are written, with published back-tests.

  • Python
  • FastAPI
  • Prophet/PyTorch
  • Postgres
  • dbt

Reference engagements illustrate how Elicon would approach a project of this kind — the problem, the architecture and the decisions. They are examples, not accounts of past client work.

THE SITUATION

The problem.

A retailer that forecasts demand by putting last year's sales next to a gut-feel column. The engagement builds a SKU-level forecasting pipeline with hierarchical reconciliation, delivers reorder suggestions into the warehouse system where buyers already work, and publishes back-tests so the team knows when to trust it.

Manual forecasting causes stockouts at peaks and write-downs in lulls.

  • ·Forecasts depend on one person's judgement and spreadsheet.
  • ·Finance and the warehouse argue about which total is real.

APPROACH

How we'd run it.

  1. 01Build a SKU-level pipeline with hierarchical reconciliation so category totals always match.
  2. 02Combine a statistical baseline with a neural model for the long tail.
  3. 03Deliver suggestions into the warehouse management system.
  4. 04Run regular back-tests and publish accuracy, flattering or not.

ARCHITECTURE

What gets built.

Feature pipeline

dbt models over sales, promotions, price changes, stockouts and calendar effects. Stockouts are masked rather than read as zero demand, or the model learns to under-forecast the SKUs that matter most.

Two-model ensemble

A statistical baseline for seasonal, well-behaved SKUs and a neural model for the intermittent long tail, selected per SKU by back-tested accuracy.

Delivery into the warehouse system

Reorder suggestions written into the warehouse system with the forecast, confidence interval and reasoning attached. Buyers accept, edit or reject where they already work.

DECISIONS

The trade-offs.

Mask stockout periods instead of treating them as demand

Zero sales during a stockout is censored data, not low demand — the most common way retail forecasting quietly fails.

TRADE-OFFLess usable history for the SKUs that stock out most.

Exclude anomalous periods rather than model them

Periods of one-off behaviour teach a model patterns that will not recur.

TRADE-OFFGaps in history and weaker long-horizon seasonality for some categories, documented so nobody rediscovers it as a bug.

Suggestions in the warehouse system, not a forecasting dashboard

Adoption is a delivery problem. A better forecast that lives somewhere buyers don't go changes nothing.

TRADE-OFFIntegration is fiddlier than a standalone app.

STACK

Typical stack.

DATA
dbtPostgresFeature pipelineStockout masking
MODELLING
ProphetPyTorchHierarchical reconciliationRolling-origin back-tests
SERVING
PythonFastAPIScheduled batch runs
DELIVERY
Warehouse system integrationConfidence intervalsAccuracy dashboard

Work of this shape usually runs as AI Product Launch →

SIMILAR PROBLEM?

Tell us what you're dealing with.

One conversation with an engineer is usually enough to tell whether this is a few weeks of work, a few months, or something you shouldn't build at all.

Talk to an engineer →