Services ยท Sales & Operations Planning
Demand forecasting
Every SKU at every site, forecast by the day for 1 to 52 weeks, with a low, most likely, and high quantity.
A forecast you can check.
Each kind of SKU is tested against several models on your own history and gets whichever forecasts it most accurately. Steady sellers, slow movers that sell a few days a month, new launches, and promoted items are each handled on their own terms, and every forecast arrives with the accuracy it earned in testing.
- By SKU, site, and order type, daily or weekly
- Low (P10), most likely (P50), and high (P90)
- Accuracy as MAPE, WAPE, or MAE against a simple baseline
- Cases, layers, and pallets converted to eaches
Four files from systems you already run.
The forecast reads flat exports from your order management system or WMS. Three are required. The fourth teaches it what your prices and promotions do.
-
Order lines
Required
At least a year of history; two years lets it learn yearly seasonality. When each order was placed, what was ordered and in what unit, the price, the customer and order type, and where it shipped from and to.
-
SKU master
Required
Every SKU on the order lines, with pack sizes for anything sold in cases, layers, or pallets, and launch and discontinue dates where you have them.
-
Site table
Required
Each shipping site's zip and time zone, so every day starts and ends in the site's local time.
-
Price and promotion plan
Optional
Past and planned prices and promotions, so the forecast learns what a promotion does and expects the next one.
Every file is checked before anything is modeled. If more than 1% of rows can't be used, or there is less than a year of history, the run stops and lists everything to fix. Smaller problems are listed as warnings.
The right model for each kind of SKU.
SKUs are grouped by how they sell. Each model is scored on the weeks it didn't see, and each group gets the model, or the blend, that did best. New SKUs borrow from similar ones until they have history of their own.
| Model | What it suits | Kind |
|---|---|---|
| Seasonal naive | The same day last week. The baseline every other model has to beat. | Baseline |
| ETS | Steady sellers with a weekly pattern. | Smoothing |
| Croston/TSB | Slow movers that sell on only a few days a month. | Intermittent |
| LightGBM | SKUs driven by price, promotions, season, and what they are bought with. One model learns across the whole catalog. | Machine learning |
Tested on your history before you rely on it.
What drives the forecast
- Day of the week and time of year
- Price and promotions, past and planned
- Customers and order type
- Where orders ship to
- Category, and what each SKU is bought with
- Launch and discontinue dates
What you receive
- The forecast with its low and high, by site and order type, in days or weeks
- Accuracy against a simple baseline, as MAPE, WAPE, or MAE
- The last 13 or 26 weeks re-forecast as they stood 1 to 4 weeks out, next to what actually sold
- Every forecast table as CSV, and as Parquet for other systems
- A handoff that sizes a Storage Strategy run to the forecast
The sample input files describe a fictional home-goods seller with three sites. They show the layout each file takes.