Melt the clock Tidy time series analysis
summary
Two new R packages tsibble and fable form the tidyverts workflow: tsibble adds an index plus a key to ordinary data frames so 46 million rows across 2,924 households become a time-series table, identifies ~20% implicit gaps, and aggregates or rolls windows (sliding, tiling, stretching) with parallel backends. fable fits scalable models (1,480 household ETS/naive fits shown), produces full predictive distributions with back‑transforms, and reports slightly better RMSE for ETS; the tools demonstrably work on large irregular datasets.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:49Tidy data workflow
- 1:20Current time series workflow
- 2:05Tidy time series workflow
- 2:52An example: electricity consumption
- 3:34Time series has its own semantics
- 4:35Contextual pretty printing
- 5:47Implicit missing value handlers
- 7:08Functional programming: map and roll
- 7:33slide
- 8:24Rolling averages
- 8:53Rolling forecast
- 9:59The first 30 days on the calendar
- 10:38mable: model() the data
- 12:24Inspect mable
- 13:18fable: forecast() the future
- 14:22Assess model performance
- 15:10Models are fundamentally scalable
- 16:05to unlock series across tsibble, mable and fable
- 16:48Joint work
related talks
22mModelling, Tidyverse Way
05mWhy tidy data matters
18mtidymodels: Adventures in Rewriting a Modeling Pipeline
20mgtsummary: Streamlining Summary Tables for Research and Regulatory Submissions
51mProgramming in the Tidyverse
29mGrammars of Data: lessons from ~20 years of the tidyverse
05mImplications of R Syntax in Intro Stats
31mAdvanced machine learning time series....