Demand Forecasting: Accuracy, Stability, Automation
summary
Market-level demand forecasts run for a 16-week horizon across 25 markets, produced daily and weekly with separate models for short and long horizons, postprocessed to correct recent errors and reconciled across temporal and product hierarchies to sometimes improve accuracy. Production pipeline uses AWS SageMaker, Databricks, Airflow with data quality checks, and planners gradually cede control as forecasts become trusted, trading some accuracy for stability over short-term accuracy so planning remains consistent and operational.
AI-written summary · about
chapters (18)
- 0:00Demand forecasting intro
- 2:17Why data-driven matters
- 3:23Accurate forecast properties
- 5:23Data prep and model building
- 6:53Forecasting time horizons
- 8:00Scaling for large markets
- 9:30Balancing accuracy and stability
- 11:22Automating forecasting
- 12:24Beyond demand to KPIs
- 14:12Forecast consolidation
- 16:36External data for forecasts
- 19:38Market-specific vs unified models
- 21:23New vs old models
- 22:15Economic indicators in forecasts
- 23:39New products in forecasting
- 24:42Forecasting new markets
- 26:29Handling Black Friday
- 27:12Conclusion and Q&A
related talks
29mProbabilistic demand forecasting at scale
35mBuilding a successful inventory planning solution with Amazon Forecast
28mUse case in demand forecasting
40mMulti-Series Forecasting at Scale with StatsForecast
18mUsing Transformer Models to Predict Demand at Scale in Production
28mHow Human Expertise and AI drive Scalable Forecasting Success at
39mProductionising Data Science at Scale
16mTransforming Inventory Forecasting with Production ML