Time series for scikit-learn people
summary
Frames forecasting as supervised ML and ships a scikit-learn compatible skits library with autoregressive, differencing and horizon transformers so you can build an X matrix, use XGBoost or linear models, and put everything in a forecaster pipeline. Demonstrates production: thousand models generating two‑hour forecasts every five minutes (thousands of forecasts per hour) served via AWS Lambda + S3 + EC2 and a Flask queue. XGBoost wins offline but linear regression is deployed for speed and monitoring shows errors vary with time.
AI-written summary · about
chapters (15)
- 0:00Time series for scikit-learn
- 1:30What is a time series
- 4:57ML to time series challenges
- 8:27Building X matrix with lags
- 11:58Adding external features
- 13:00Missing lags in deployment
- 14:23Differencing time series
- 15:27Introducing skits
- 17:37Recursive vs horizon prediction
- 20:31Forecasting pipelines
- 21:58Multi-output regression
- 23:02City bike data training
- 24:47Real-world forecast evaluation
- 26:12Summary and skits
- 26:41Q&A
related talks
26mTime Series Analysis in Python: Easy Tools for Scientific Insight
30msktime: python toolbox for time series- next-generation AI
19mForecasting for Everyone - Foundation Models and the Future of Time Series Prediction
42mRobust Automated Forecasting in Python and R
42mRobust Automated Forecasting in Python and R
30mReal-Time ML services: From ideation to production
28mEfficient Time-Series Forecasting with Thousands of Local Models on Databricks
22mScalable AutoML for Time Series Forecasting