Forecasting at Scale
summary
Describes concrete techniques: compute ~25 statistical features per series, run PCA then cluster to group similar behavior, and pick representative series per cluster to design features and models. Shows a working R app that runs this on 250 series of monthly natural gas consumption (five categories), uses backtesting orchestration for rolling evaluation, and compares models like AR/ETS in an autoML-style race. Claims this reduces effort from per-series modeling to cluster-driven transfer and notes tradeoffs in accuracy and compute.
AI-written summary · about
chapters (13)
- 0:00Forecasting at scale intro
- 1:02Talk agenda
- 3:14Time series basics
- 7:10Traditional forecasting limits
- 7:53Defining forecasting at scale
- 9:20Scalable forecasting architecture
- 10:25Back testing as cross-validation
- 12:32Automated model racing
- 15:45Feature engineering at scale
- 17:33Time series clustering
- 23:13Live clustering demo
- 30:56Key takeaways
- 33:10Question and answer session
related talks
40mMulti-Series Forecasting at Scale with StatsForecast
42mRobust Automated Forecasting in Python and R
01h 26mApplying Statistical Modeling & Machine Learning to Perform Time-Series Forecasting
28mEfficient Time-Series Forecasting with Thousands of Local Models on Databricks
42mRobust Automated Forecasting in Python and R
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
29mProbabilistic demand forecasting at scale
40mDeep Learning and Time Series Forecasting for Smarter Energy