[ENH] Interface statsmodels MSTL - transformer #5125
Conversation
luca-miniati
requested review from
achieveordie,
benHeid,
fkiraly and
yarnabrina
as code owners
There was a problem hiding this comment.
Nice, great!
A question, not necessarily a change request:
In the statsmodels example here
https://www.statsmodels.org/devel/generated/statsmodels.tsa.seasonal.MSTL.html
it seems that MSTL returns have as many seasonal components (named Seasonal_[period] in the plot) as there are entries to the iterable periods. Whereas, this transformer seems to add the seasonal components all up?
The result must allow to access these somehow, since the plots displays the components separately. Although there's no clear documentation in statsmodesl that I can find...
|
Another comment, there is also this PR, seems same name? It makes sense to have one PR with the transformer, and one (the other one?) with a forecaster - where the forecaster makes forecasts for the seasonal components and the trend component, with defaults being |
I've split the seasonal components as shown in the |
Yes, I had moved the file and forgot to add the "new" file. Should be there now under |
There was a problem hiding this comment.
Looks great!
Blocking change request: can you explain to the user in the docstring what happens in transform (and/or fit)?
Non-blocking change requests:
- stray newline in API reference
- should there be an inverse transform? That should just sum up the three components back to one, as far as I see.
fkiraly
changed the title
…encies * origin/main: [DOC] speed-up tutorial notebooks - deep learning classifiers (sktime#5169) [ENH] fixture names in probability distribution tests (sktime#5159) [ENH] test for specification conformance of tag register (sktime#5170) &SmirnGregHM [BUG] ensure forecasting tuners do not vectorize over columns (variables) (sktime#5145) [ENH] VMD (variational mode decomposition) transformer based on `vmdpy` (sktime#5129) [ENH] Interface statsmodels MSTL - transformer (sktime#5125) [ENH] add tag for inexact `inverse_transform`-s (sktime#5166) [ENH] refactor and add conditional execution to `numba` based distance tests (sktime#5141) [MNT] move fixtures in `test_dropna` to `pytest` fixtures (sktime#5153) [BUG] prevent exception in `PyODAnnotator.get_test_params` (sktime#5151) [MNT] move fixtures in `test_reduce_global` to `pytest` fixtures (sktime#5157) [MNT] fix dependency isolation of `DateTimeFeatures` tests (sktime#5154) [MNT] lower dep bound compatibility patch - `binom_test` (sktime#5152) [MNT] test forecastingdata downloads only on a small random subset (sktime#5146) [ENH] widen scope of change-conditional test execution (sktime#5147) [DOC] update forecasting extension template on `predict_proba` (sktime#5138)
What does this implement/fix? Explain your changes.
Interfaces statsmodels MSTL, replaces #4970
For new estimators
dependency isolation, see the estimator dependencies guide.