TimeSeriesImputerTransform

class TimeSeriesImputerTransform(in_column: str = 'target', strategy: str = ImputerMode.zero, window: int = - 1, seasonality: int = 1, default_value: Optional[float] = None)

Bases: etna.transforms.base.PerSegmentWrapper

Transform to fill NaNs in series of a given dataframe.

• It is assumed that given series begins with first non NaN value.

• This transform can’t fill NaNs in the future, only on train data.

• This transform can’t fill NaNs if all values are NaNs. In this case exception is raised.

Warning

This transform can suffer from look-ahead bias in ‘mean’ mode. For transforming data at some timestamp it uses information from the whole train part.

Create instance of TimeSeriesImputerTransform.

Parameters

• in_column (str) – name of processed column

• strategy (str) –

  filling value in missing timestamps:

  • If “zero”, then replace missing dates with zeros

  • If “mean”, then replace missing dates using the mean in fit stage.

  • If “running_mean” then replace missing dates using mean of subset of data

  • If “forward_fill” then replace missing dates using last existing value

  • If “seasonal” then replace missing dates using seasonal moving average

• window (int) –

  In case of moving average and seasonality.

  • If window=-1 all previous dates are taken in account

  • Otherwise only window previous dates

• seasonality (int) – the length of the seasonality

• default_value (Optional[float]) – value which will be used to impute the NaNs left after applying the imputer with the chosen strategy

Raises

ValueError: – if incorrect strategy given

Inherited-members

Methods

fit(df)	Fit transform on each segment.
fit_transform(df)	May be reimplemented.
inverse_transform(df)	Apply inverse_transform to each segment.
transform(df)	Apply transform to each segment separately.