Time-aware feature generation
Specify how feature engineering should be applied to external features (see External features).
For each external feature, you can define lag values and aggregates computed over one or more rolling windows.
These engineered features will be used by algorithms that support external features.
None of the selected algorithms can use features generated here.
Time-shifted features
Dataiku generates lagged features (or "negative shifts") by using data from the past.
Feature
Role
From forecast origin
A feature value from a fixed time step before the forecast origin (when the forecast
is made).
Example: A forecast origin shift of -2 uses the same value from 2 steps before the
forecast origin, for all the predicted steps in the horizon.
From forecasted point
A feature value from a time step that "slides" relative to each point in the
forecast horizon.
Example: For a forecast from t+1 to t+2, a forecasted point shift of -7 uses the
value at t-6 for the t+1 forecast and t-5 for the t+2 forecast.
In Auto mode, Dataiku explores various forecasted point shifts within the specified
range.
Auto handling
Aggregate features
Compute statistics like average, std dev, min, max, or most frequent over a time window (e.g., 35 to 0 days) before forecast origin or forecasted point.
Aggregates before the forecast origin stay the same for all forecast steps; those before each forecasted point update individually for every step in the forecast horizon.