CONCEPT Cited by 1 source
SKUs as time-series unit¶
Definition¶
In e-commerce demand-forecasting systems, each SKU (Stock-Keeping Unit) is treated as an independent time-series unit: the forecast is computed per SKU, and the forecasting model sees each SKU's sales history as its own series.
The natural keying question is what defines a unique SKU for forecasting purposes:
- Article ID alone — too coarse if the same article is sold by multiple merchants with different demand dynamics.
- (Article ID, Merchant ID) — finer; captures per-merchant demand differences.
- (Article ID, Merchant ID, Size, Colour) — finest; captures per-variant demand differences.
Each step finer multiplies the number of series by the number of distinct values of the added dimension — a cardinality explosion tradeoff with implications for training cost, sparse-series handling, and hierarchical-aggregation choices.
Canonical instance (Zalando ZEOS)¶
Zalando treats each (article_id, merchant_id) pair as
a unique time-series entity:
"unique identifier per time-series: we treat each combination of (article_id, merchant_id) as a unique entity."
And reports the final-granularity catalogue scale:
"Our weekly forecasting pipeline processes 3 years of historical data for 5 million SKUs (size and colour) using a sliding window approach."
So the system is keyed by (article_id, merchant_id) for
the time-series granularity, and downstream the forecast
horizon is 12 weeks per
(article_id, merchant_id, week). Size + colour variants
are counted in the 5-million-SKU figure but appear to be
modelled at a more aggregated level than the fundamental
time-series unit — the post is not fully explicit on the
size/colour → series mapping.
Why the merchant dimension matters¶
Zalando is a marketplace: the same article can be sold by multiple merchants, with different pricing, logistics, and demand patterns. Forecasting at the (article, merchant) level gives merchants individualised replenishment recommendations that reflect their specific selling velocity and inventory history, even for articles shared with other merchants.