Regular Contributors Meetings

2021-01-27

Attendees: Michal Chromčák, Pavel Křížek, Jan Therhaag

Usage in non-daily data
- @Pavel to write issues
- The current implementation of HolidayTransformer does not support different than daily frequency - let's add it
  - Aggregation of daily data to e.g. monthly data
    - Resample to daily, add holidays, aggregate back
  - Minute data
    - Resample to daily, add holidays, fill back
- Do not restrict the SeasonalityTransformer just to the daily data
  - Similar to HolidayTransformer approach
  - Introduce features for hours etc.
During model selection, models are always re-fit. In some use-cases, this is very costly and just update of data vs. refit could help
- @Pavel to write the exploratory issue
Minimal data length does not fail silently for SklearnWrapper in the model selection
- @Michal to write the issue

Attendees: Michal Chromčák, Pavel Křížek, Markus Löning

sktime HCrystalBallWrapper
- Current state in PR 485
- Minimal working version (no support for in-sample predictions, prediction intervals, or index types other than date-time)
- Next steps
  - Merge the pull request with this minimal form
  - If any feedback comes, react respectively
NeuralProphet
- Promising new model built on top of PyTorch, with a very similar interface to the prophet
- Already Opened ISSUE 515 at sktime
- Next steps
  - TBD
CyclicBoostingRegressor
- No code available - would need to implement it by the paper
- Next steps
  - If to implement - create it in a different library and make a wrapper for it here
Detrend and deseasonalize transformers
- Problem: Tree-based models do not cope well with unseen range, which is often a case of series with a strong trend.
- Next steps
  - Have a look at sktime's implementation of detrenders and whether to re-use it or do a custom implementation