Possible to see transformed time series data?

[stikket]

Hi everyone, great package.

Is it possible to see the transformed version of training data for timeseries? Is this feature available already in the package?

[winedarksea]

I assume you are looking for what the data looks like after the preprocessing transforms are applied, as the underlying model would see the data?

Ensembles can mess things up a bit because they mix together separate transformer+model groups, but generally:

import json
from autots import GeneralTransformer, AutoTS

model = AutoTS(...your_args...).fit()

# Ensembles don't have one overall transformer, so we can't plot that
# but for horizontal with no style 1 ensembles we can still view the best non ensemble model transformer
if model.best_model_ensemble == 2:
    params = json.loads(model.best_model_non_horizontal['TransformationParameters'].iloc[0])
else:
    params = model.best_model_transformation_params
    params = model.best_model_transformation_params
# if this is an empty dict, this means that best model and best non horizontal are both an Ensemble
# you could loop through the Ensemble params and plot each transformer, in that case, however, not shown
print(f"Transformer params are {params}")

transformer = GeneralTransformer(**params)
transformed_df = transformer.fit_transform(model.df_wide_numeric)

column = model.df_wide_numeric.columns[0]
plot_df = model.df_wide_numeric[column].to_frame().merge(transformed_df[column].to_frame().rename(columns=lambda x: "transformed_" + str(x)[0:50]), right_index=True, left_index=True)
plot_df.plot() # scale is often very different between two, because scaling
plot_df.iloc[:, -1].plot() # just the transformed
# plt.show()

[stikket]

Thank you for giving a very concise answer.

A few follow up question:
If I have a single column to forecast and have multiple transforms, the transformed_df would still have a single column? I was assuming it would add columns (or is it the time series model itself that creates additional features? i.e. ARIMA and it would just use the transformed column as input?)

[winedarksea]

So there are two different things which are both important:

1. The preprocessing transformers. These work in 'layers' to add and remove various aspects to the data around the primary model
2. The primary model, which in the case of ~Regression style models and some others, will have an X and a Y feature set which do have columns of features created from the input series.
   My previous answer focused on point 1.

So, details:

• Most transformers don't add columns. They instead work in layers adding and removing but mostly keeping the original data shape at the end. The exceptions are CenterSplit and anything with a macro_micro arg which do double the number of columns by splitting the data
• If you are looking for models that provide an X and a Y feature set for other machine learning, the easiest to extract that from are the DatepartRegression, MultivariateRegression, and WindowRegression. You can optimize these with AutoTS, then use both the transformers and model parameters to extract a feature set. I can provide code for this if you need help with that.
• PreprocessingRegression (still kinda beta) is a model that mixes the preprocessing as the X feature set, so it is an exception.