Learning project on public data.
Can a simple forecast and plan stand up to a messy demand history?
I used the Cups4U case from MIT to forecast 12 months of paper cup demand, test the error honestly and compare a chase plan with a level plan.
Rows of order dataAug 1996 to Aug 20064,21612 SKUs, 18 customers
Last 12 monthsCases ordered5,247,496+2.0% on the year before
Test WAPESeasonal exponential smoothing12.5%Same month last year: 12.1%
Level outputCases a month, next 12403,619No shortage with zero start stock
What I built
- Measured channel trends, the lost customer, the price rise and promotion lifts from the raw rows.
- Picked a forecast method on one year and tested it on the next, unseen year.
- Built chase and level plans with an inventory balance and editable cost rates.
- Added formula cards with worked examples on every screen.
What the numbers show
The chosen seasonal method scored a test WAPE of 12.5%. The plain same month last year rule scored 12.1%, so a simple baseline was hard to beat in a year with a lost customer and a price rise. The plan costs depend on assumed rates, which the app lets you change.