Workspace/Overview
Historical replayFF
DEMAND & REPLENISHMENT

Demand & replenishment plan.

Every order backed by data.

Loading forecast dates7 days
Order today
—
Pairs with an order to place
High waste risk
—
Simulated shelf life vs planned stock
In view
—
Store × category
v3 backtest WAPE
—
Mean of 4 retrained windows (naive: 19.9%)

Next 7 days demand

Pick a row in the replenishment table to see its forecast

LightGBM v3 · p50 with p10–p90
—7-day forecast total (p50)
Forecast (p50)80% range (p10–p90)
Loading forecast…
Sales in source-data unitsHistorical replay · not live data

Replenishment

—
Click a row to update the chart and assistant
Stock is simulated
Store / category7-day forecastSimulated stockOrder for next deliveryActionWaste riskView
Loading recommendations…
—For planning only · no purchase order is created
BEHIND THE NUMBERS

Where do the numbers come from?

A demand-planning prototype built on historical sales data.

01

Real historical sales

Favorita retail data for 54 stores and 13 food categories, through 2017-08-15; forecasts cover 08-16 to 08-22. Categories are product families, not SKUs, and sales keep their source units.

02

LightGBM v3 forecast, backtested

A direct 7-day model using only sales known on the Wednesday order date, plus planned promotions, store-matched holidays and paydays. Retrained on 5 historical windows: 11.7% mean WAPE vs 19.9% for last-week-same-day. The p10–p90 band covered 77% of actuals (target 80%).

03

Transparent simulated operations

Stock, shelf life, lead time and MOQ are simulated. Orders cover demand until the next delivery plus a per-category buffer learned in the backtest (the policy behind the 51% cost saving). Promotion what-ifs re-score the model with a changed promotion plan. Recommendations are not turned into purchase orders.

04

An assistant that shows its evidence

An LLM agent answers through tools over the v3 artifacts; numbers are checked against tool results. If the LLM is unavailable, a labelled template fallback is used. Tool calls and raw results can be expanded under each answer.