Probabilistic demand forecasting at scale
summary
Shows how probabilistic forecasts replace blind point estimates by outputting quantiles or learned distribution parameters, trained with pinball loss or likelihood and evaluated with CRPS and cross entropy (WAPE for business impact). Implements deep models like Temporal Fusion Transformer, scales training on GPU clusters and stores artifacts on S3 / SageMaker, and runs inference via SageMaker batch transform. A concrete claim: forecasting about 150,000 time series in 45 minutes for daily jobs, enabling stochastic inventory and capacity planning.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:11Demand forecasting problem
- 4:02Covariates and features
- 6:33Limitations of deterministic forecasting
- 11:24Uncertainty with probabilistic modeling
- 13:37Probabilistic forecasting methods
- 16:28Distribution parameters vs quantiles
- 19:36Evaluating probabilistic forecasts
- 22:04Libraries and tools
- 22:47End-to-end workflow
- 25:52Q&A
related talks
28mDemand Forecasting: Accuracy, Stability, Automation
40mMulti-Series Forecasting at Scale with StatsForecast
18mUsing Transformer Models to Predict Demand at Scale in Production
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
35mBuilding a successful inventory planning solution with Amazon Forecast
54mTime Series forecasting with NumPyro
42mRobust Automated Forecasting in Python and R
39mProductionising Data Science at Scale