Vhona AI
Talk to us

Sample

Axum Home

Orders come in, get cleaned, and a model forecasts the next two days for each product and region.

Prototype. Not production data.

  1. 1Skynamo order export
  2. 2Clean and aggregate by SKU, region, and day
  3. 3Prophet model per SKU and region
  4. 42 day forecast and a confidence range
  5. 5Next: an Azure ML pipeline on a schedule, feeding dispatch planning

Glass cleaner lemon, Johannesburg

GC210 · Gauteng, Johannesburg zone. Last 30 days actual, next 2 days forecast, with the confidence range.

Needs attention

Low confidence, high volume

Biggest volume forecasts where the model’s uncertainty is widest.

SKURegionForecastConfidence

Next 2 days

SKU Region Day 1 forecast Day 2 forecast Confidence

Forecasted units, day 1, by region, top 15

Forecasted units, day 1, by SKU

Combos not forecasted yet

SKU and region pairs with under 30 days of order history. Not enough to learn a weekly pattern from. Ranked by volume, since these are the ones worth watching as more data comes in.

SKURegionDays seenTotal units (3mo)

What was thrown out

Cleanup applied to the raw export.

  • 1,141 cancelled orders excluded. Never shipped, not real demand.
  • “kivo” codes reclassified as a customer code, removed from the region breakdown.
  • 299+ one off region codes folded into “other” (typos, stray notes, miskeyed order numbers).
  • Stray far future schedule dates (a handful of orders dated months ahead) capped out of the training window. They were distorting the weekly pattern.
  • 229 real SKU columns reshaped from the wide order sheet into one row per day, region, and SKU.

Forecasts are generated by a Prophet model trained per SKU and region on about 3 months of Skynamo order history. Confidence reflects the width of the model’s prediction interval relative to the forecast size. A narrow range is high confidence. Open any row in the forecast to see that product.

Also in Demos: Karoo Kitchen