[BUG] Why is the inverse transformation of the target not applied?

[ngupta23]

Discussed in #2706

^{Originally posted by acartro July 4, 2022}
Good morning,

I am doing a time series experiment with the new 3.0 library. I have found that when I set the setup to do a logarithmic transformation of the target (transform_target='log') when making the forecasts, it applies the inverse transformation.

Whit target transformation

from pycaret.time_series import *
from pycaret.datasets import get_data
y = get_data("airline")

#### Setup experiment ----
exp = TSForecastingExperiment()
exp.setup(data=y, fh=12, session_id=42, transform_target='log')

##### List the available models ----
exp.models()

#### Create a model from available models ----
model = exp.create_model("arima")

### plot model
exp.plot_model(model )

Whit target transformation and impute missings

But when I set the setup to do an imputation of the missings and then the logarithmic transformation (transform_target='log', numeric_imputation_target='drift'), when it makes the forecast, it doesn't apply the inverse transformation.

#### Setup experiment ----
exp = TSForecastingExperiment()
exp.setup(data=y, fh=12, session_id=42, transform_target='log',
          numeric_imputation_target='drift')

##### List the available models ----
exp.models()

#### Create a model from available models ----
model = exp.create_model("arima")

### plot model
exp.plot_model(model )

Does anyone know why this is happening or am I just doing the forecast wrong?

[ngupta23]

Also see: #2724

[ngupta23]

I am able to reproduce this. Will need to dig deeper and get back to you.

https://gist.github.com/ngupta23/981f287598761e69767a0c17d4a465b5

[ngupta23]

@acartro @thebetauser This seems to be a bug in sktime. I have opened an issue in their repository and will wait for their response

sktime/sktime#3084

[ngupta23]

The issue is fixed in sktime (unreleased version). This should be fixed with the next release of sktime and pycaret.