Mastering Demand Forecasting: Lessons from Europe's Largest Retailer
summary
Retail demand forecasting runs on a feature‑based GBM pipeline using LightGBM, training one model per item‑warehouse and running hundreds of thousands of models retrained weekly. Features include calendar, holiday, promotion, store‑physics and lag terms; two tweaks — feature contribution and sample weighting — cut aggregated holiday MAPE by about 6% and a dedicated holiday uplift model with a decompose‑recompose step further reduces error from 100% baseline to roughly 75%, deployed into automated replenishment.
AI-written summary · about
chapters (12)
- 0:00Demand forecasting intro
- 1:33Manual forecasting failures
- 3:38Supply chain forecasting
- 6:48Time series as tabular regression
- 10:58Base model performance
- 13:26Holidays break models
- 17:57Holiday predictions with uplift
- 20:26Holiday forecast accuracy
- 21:52Specialized forecasting
- 23:38Postprocessing and future vision
- 24:20New products and data drift
- 26:46Q&A
related talks
30mAI in Retail - Advancing Promo Demand Forecasting
28mUse case in demand forecasting
23mMachine Learning with a Deep Dive into Automated Forecasting System
28mDemand Forecasting: Accuracy, Stability, Automation
28mHow Human Expertise and AI drive Scalable Forecasting Success at
59mPredictability Check: Demand Forecasting
35mBuilding a successful inventory planning solution with Amazon Forecast
09mDemand Forecasting in Supply Chain Using AI Basis Machine Learning