Lag features

Window features

• How to choose window size & statistics?
• Rolling windows or Expending windows?
• Add weights to the windows

Rolling window

Edge case 1: drop NaN

Edge case 1

Edge case 2: smaller window size

Edge case 2

How to pick window size?

• Huristic - set the window size to be the same as the seasonal period(s)
• Trail & Error - try multiple window sizes and use feature selection method to decide the best

How to pick statistics?

• mean
• std
• count, sum, min, max, corr, cov, skew, kurt,…

Implementation

Pandas

df['y'].rolling(window=3).agg(["mean","min"]).shift(periods=1)
df['y'].rolling(window=3, min_periods=1).agg(["mean","min"]).shift(periods=1)

Feature-engine

from feature_engine.timeseries.forecasting import WindowFeatures

transformer = WindowFeatures(
    variables = ['y'],
    functions = ['mean', 'std'],    # statistic
    window = [1, 3, 6],    # window size
    freq = '1MS')
transformer.fit_transform(df)

sktime