Forecast demand
—units
VN2 forecastNext 90 days
DEMAND → DECISIONS
Inventory overview.
Know what to order. See what runs out next.
Opening inventory
—units
Scenario35 days of historical cover
Stockout risk
—items
— urgentWithin selected horizon
Suggested replenishment
—units
Policy estimate— purchase suggestions
FORECAST QUALITY → INVENTORY DECISIONS
Forecast workbench
Loading saved pipeline models…
Changing the model recalculates demand, replenishment and stockout dates. Historical evaluation is separate from the future forecast; a lower error is not a guarantee of fewer stockouts.
Holdout WAPE ↓—Absolute error / actual demand
Bias-adjusted WAPE ↓—WAPE + absolute percentage bias
Signed bias—Negative = underforecast
Scored item-months—Same coverage for every model
Compare models · sorted by CV bias-adjusted WAPE, not holdout
CV is used to tune the blends; it is not an independent test. The historical holdout is reported separately. The archived default is retained rather than replaced with the best-looking holdout model.
| Model / family | CV BA-WAPE ↓ | Holdout WAPE ↓ | Holdout BA-WAPE ↓ | Holdout bias |
|---|
Ensemble weights, metrics and provenance
BA-WAPE is a bias penalty, not a bias-corrected forecast. Errors are summed over individual item-months before division; zero-demand scopes have no defined percentage score. H1 means one month after the forecast cutoff, not the first calendar month of the year. The CV-derived blend uses inverse CV BA-WAPE with 25% equal-weight shrinkage; it does not tune on holdout data.
Demand & forecast
Observed monthly demand and the 12-month forecast.
Actual demand Ensemble_Horizon
Historical dates preserved12-month forecast · — units
Inventory projection
Baseline vs one suggested replenishment order.
No new order With order Buffer
Unfilled units · baseline—
With suggested order—
YOUR REPLENISHMENT WORKLIST
Inventory items —
Stockout dates estimate the first day demand cannot be filled. Select an item to inspect its curves.
| Product / location | Status | On hand ⓘ | Inbound | Est. stockout | Order by | Buy units | Arrival / lead | Edit assumptions |
|---|
No matching inventory items
Try another location, product or risk filter.
—
1 / 1
Data sources & planning methodology Real demand. Explicit assumptions. No invented accuracy claims.
VN2 demand data
History: input/vn2_monthly_complete.csv
Forecast: run_h12_cv1/inventory_demand_quantiles.csv, column Ensemble_Horizon.
Inventory is a what-if scenario
Opening stock defaults to the last three observed months’ daily average × the selected cover days, rounded up. Inbound defaults to zero. Edits stay in this browser session only.
Monthly forecasts are spread evenly over their actual calendar days. Receipts arrive before that day’s demand. Unfilled demand is counted as lost sales; stock never goes negative. A new order arrives after the stated lead time.
A transparent replenishment policy
Suggested units cover lead time + review period + buffer days, less opening stock and inbound arriving within that window, rounded up. This is a policy simulation, not a globally optimal purchase plan. A late order cannot recover earlier lost sales.
Dates are approximate. There are no live stock, supplier, cost, capacity, MOQ or transfer records here. No dates or demand are extrapolated beyond the source forecast. Portfolio totals can hide individual shortages.
The model switcher uses saved future_forecasts.csv, CV and holdout predictions from the same VN2 run. Selecting a model changes the forecast and inventory scenario, not observed demand. BA-WAPE penalizes bias; it does not automatically bias-correct forecasts.