CONCEPT Cited by 1 source
Sliding-window training¶
Definition¶
Sliding-window training is a time-series training strategy where the training dataset is a fixed-length window that slides forward in time with each retraining cycle. Instead of accumulating all historical data forever, the window drops old data as new data enters — bounding training cost, keeping the model responsive to recent dynamics, and avoiding overweighting ancient history.
Why sliding-window¶
- Bounded training cost. Training time scales with window length × number of entities. Fixing the window length makes training cost flat over time even as history grows.
- Model recency. Recent patterns (new products, evolving customer behaviour, new competitors) get representative weight; very old data is less informative.
- Captures seasonality. Window length is chosen to cover enough seasonal cycles (years for annual seasonality) while not over-representing ancient regimes.
- Avoids concept drift from distant history. 10-year-old sales patterns may be actively misleading for a fast-moving catalogue.
Canonical instance (Zalando ZEOS)¶
- Window length: 2.5 years (selected from 3 years of available history).
- Rationale verbatim: "we use a 2.5-year timeframe to enable the model to capture seasonal patterns without overemphasising older historical performance."
- Entities: 5 million SKUs at size + colour granularity.
- Cadence: weekly retraining.
- Producer: systems/zeos-demand-forecaster via Nixtla MLForecast on LightGBM.
Seen in¶
Related¶
- concepts/probabilistic-demand-forecast — Zalando's downstream output.
- systems/zeos-demand-forecaster · systems/mlforecast-nixtla
- companies/zalando