functime supports parallelized time-series preprocessing using Polars. All
functime preprocessors take a panel DataFrame as a input and transform each time-series locally (i.e. time-series by time-series as a parallelized group_by operation).
Time-series transformations are commonly used to stabilize the time-series (e.g.
boxcox for variance stabilzation) or make the time-series stationary through first differences or detrending. Some transformations are also invertible, such as
detrend, which is useful for converting the forecast of a transformed time-series back to the original scale.
Check out the API reference for details.
Apply k-order differences. This transform is invertible.
Apply k-order differences shifted by
sp periods. This transform is invertible.
Removes linear trend for each time-series. This transform is invertible.
Removes mean trend for each time-series. This transform is invertible.
Applies optimized Box-Cox transform for each time-series. This transform is invertible.
Applies optimized Yeo-Johnson transform for each time-series. This transform is invertible.
Standardizes each time-series with subtracting mean and dividing by the standard deviation. This transform is invertible.
Given a list of window sizes, applies rolling statistics for each time-series across each column. This transform is not invertible. Currently supports the following statistics:
mlm (max less min),
cv (coefficient of variation).