01
How many orders next week?
1,142 a week on average. Some weeks 46, some 2,979.
02
I bet on machine learning.
XGBoost lost, even to “same as last week”.
03
Simple structure won.
With 86 weeks of history, a classic SARIMA model beat them all: 18.5% error.
04
Promised 23 days. Delivered in 12.
Olist padded its estimates. 92.1% arrived on time.
05
881 sellers. One real problem.
One seller was late on 32.1% of orders.
06
With more data, ML wins.
In the M5 contest on Walmart data, all top 50 teams used LightGBM.
Sources: M5 competition, International Journal of Forecasting ↗
07
Next: find the crossover.
Rerun as history grows, and see when ML takes over.