API Reference / preprocessing /

timeseries/

Turning a time series into features a model can learn from.

A forecasting model needs the past available as columns: what the value was some steps ago, and how much it has changed since. These transformers shift the selected columns in place, which is what lets an ordinary regressor be applied to sequential data.

Transformers

  • LagTransformer: Replaces a column with its value lag steps away, so
each row carries history instead of only the present.
  • DifferenceTransformer: Replaces it with the change over lag steps.
Differencing also removes a trend, which is what makes a drifting series stationary enough to model.

Notes

Mind the sign: lag=-1 looks backwards, at the previous value, and is almost always what you want. A positive lag looks forward, feeding the model a value from the future — which trains beautifully and predicts nothing. Both transformers default to lag=-1.

Shifting leaves boundary rows with no value to draw on. By default those are filled with NaN and the row count is preserved; pass fill_with_missing=False to drop them instead.

Evaluate with TimeSeriesSplit. Lagged features make a shuffled split leak outright: a training row can hold the very value a test row is being asked to predict.

python 
>>> import numpy as np
>>> from tuiml.preprocessing.timeseries import LagTransformer
>>> X = np.arange(5, dtype=float).reshape(-1, 1)
>>> LagTransformer(lag=-1, columns=[0], fill_with_missing=False).fit_transform(X).ravel().tolist()
[0.0, 1.0, 2.0, 3.0]

See Also

timeseries

Forecasting models for this data.

Modules

delta

Time series delta (difference) transformer.

1 class · 0 functions
lag

Time series lag (translate) transformer.

1 class · 0 functions