[BUG] MSTL documentation is misleading after 0.24.0 release

[ArturDragunov]

Describe the bug

Up until the recent version, MSTL had a note that it can't inverse. The same note is still present in STLTransformer: STLTransformer can not inverse_transform on indices not seen in fit(). This means that for pipelining, the Deseasonalizer or Detrender must be used instead of STLTransformer. This creates a misleading situation for the users who could have assumed that MSTL can inverse_transform.
To Reproduce

from sktime.datasets import load_airline
import pandas as pd
from sktime.forecasting.naive import NaiveForecaster
from sktime.forecasting.compose import TransformedTargetForecaster
from sktime.transformations.series.detrend.mstl import MSTL
from sktime.transformations.series.impute import Imputer
from sktime.forecasting.model_evaluation import evaluate
from sktime.forecasting.model_selection import ExpandingWindowSplitter


y = load_airline()
y.index = pd.PeriodIndex(y.index, freq="d")
data = pd.DataFrame(y)

for i in range(1, 4):
    data['Lag_' + str(i)] = data[y.name].shift(i)
data.dropna(inplace=True)
    
X = data.drop(columns=y.name)
y = data[y.name]

pipe = TransformedTargetForecaster(steps=[
    ("imputer", Imputer(method="mean")),
    ("deseasonalizer", MSTL(periods=[7])),
    ("forecaster", NaiveForecaster(strategy="drift")),
])

cv = ExpandingWindowSplitter(initial_window=15, step_length=6, fh=[1, 2, 3])
results = evaluate(forecaster=pipe, X=X, y=y, cv=cv) #, error_score='raise')
results

FitFailedWarning:
In evaluate, fitting of forecaster TransformedTargetForecaster failed,
you can set error_score='raise' in evaluate to see
the exception message. Fit failed for len(y_train)=15.
The score will be set to nan.
Failed forecaster with parameters: TransformedTargetForecaster(steps=[('imputer', Imputer(method='mean')),
('deseasonalizer', MSTL(periods=[7])),
('forecaster',
NaiveForecaster(strategy='drift'))]).

result = _evaluate_window(

test_MeanAbsolutePercentageError	fit_time	pred_time	len_train_window	cutoff
NaN	0.018867	0.009660	15	NaT
NaN	0.015766	0.007032	21	NaT
NaN	0.016762	0.010788	27	NaT
NaN	0.014447	0.007028	33	NaT

Expected behavior

Ideally, MSTL should have an inverse_transform feature available. Otherwise, the documentation description should change.

Versions
sktime 0.24.0

[fkiraly]

Hm, both STL and MSTL should be able to inverse transform in theory.

I know the feature was added for STL in one of the recent versions so the STLTransformer docstring should be updated at least.

[ArturDragunov]

@fkiraly I have a small typo in the issue. I was playing with different versions and forgot to change from 0.23.0 to 0.24.0. The error logs have changed but the output is still the same that MSTL can not inverse_transform on indices not seen in fit(). I will see the changes in the source code for STLTransformer. Will experiment with it.

FitFailedWarning:
In evaluate, fitting of forecaster TransformedTargetForecaster failed,
you can set error_score='raise' in evaluate to see
the exception message.
Fit failed for the 0-th data split, on training data y_train with
cutoff NaT, and len(y_train)=15.
The score will be set to nan.
Failed forecaster with parameters: TransformedTargetForecaster(steps=[('imputer', Imputer(method='mean')),
('deseasonalizer', MSTL(periods=[7])),
('forecaster',
NaiveForecaster(strategy='drift'))]).

System:
python: 3.11.4 (main, Jul 5 2023, 14:15:25) [GCC 11.2.0]
machine: Linux-6.2.0-34-generic-x86_64-with-glibc2.37

Python dependencies:
pip: 23.1.2
sktime: 0.24.0
sklearn: 1.3.0
skbase: 0.6.0
numpy: 1.25.2
scipy: 1.11.1
pandas: 2.0.3
matplotlib: 3.7.1
joblib: 1.3.0
numba: None
statsmodels: 0.14.0
pmdarima: None
statsforecast: None

[fkiraly]

I updated the docstring for #5455.

I looked at MSTL, my intention is just to replicate the STL inverse_transform, then it can be used in a transformer pipeline (as a decomposition transformer). The docstring and usage would then be the same.

[fkiraly]

Update for MSTL is here: #5457

I also noticed that the decomposition does not give individual components, I opened an issue here #5458 to track that we should add this as an option for use in a transformed target forecasting pipeline.