The first part argued that hierarchical forecasting in production is a representation problem, not a modeling one — and that the right architecture for luxury demand at SKU level is a single global LightGBM model with the hierarchy living in the feature space, not in probabilistic pooling.
- category, - sub-category, - brand, - price tier, - pack size, - seasonality, - lifecycle stage, - channel, - geography, - launch timing, - and commercial positioning.
The model therefore learned demand behaviour from the broader structure of the catalogue.
- "premium gifting behaviour," - "mid-price seasonal products," - "holiday luxury spikes," - or "slow-moving high-margin products."
About the author
Cyril Noirot
Lead Data Scientist
Freelance data scientist. I design and ship decision systems — forecasting, pricing, marketing measurement, optimization.