How I used Polars to build built functime, a next gen ML forecasting library
summary
Claims a concrete speedup: feature engineering runs 1-2 orders of magnitude faster using Polars; batch inference and scaling happen without PySpark. Presents an implemented library, functime, built on Polars with global-forecasting (single-model) strategy and gradient-boosted models, plus feature-extraction, backtesting, ranking diagnostics and conformal intervals. Demo shows pipeline APIs, train/test by horizon, plotting and ranking of worst series, and reports that these approaches materially reduce end-to-end forecasting cost.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:31Forecasting needs new approach
- 4:22Introducing functime
- 5:29Polars powers functime
- 8:00Global forecasting models
- 10:46Live demo of API
- 16:47Forecasting workflow
- 18:33Diagnostics and evaluation
- 20:16Building models
- 22:24Comparing forecast models
- 23:49Advanced features
- 25:34Resources and community
- 26:40Q&A
related talks
31mfunctime: a next generation ML forecasting library powered by Polars
21mPolars: A highly optimized dataframe library
45mUncertainty estimation at scale with functime
31mPolars, the End of Spark?
40mMulti-Series Forecasting at Scale with StatsForecast
36mBuild prod ML systems with feature /training /inference pipelines
24mPlatform-First Forecasting: Engineering Scalable ML Systems
39mDesigning for high-performance execution with Arrow and Polars