Moving Window Transforms
Moving window transforms operate on the portion of an input variable that is within a “moving window.” They all take as input at least a variable, an integer window size, an alignment constant, and a threshold (some of them take additional input arguments, specified below; the threshold is placed after all other arguments). The size automatically shrinks as boundaries and breaks are approached. The alignment constant must be one of $lag, $center, or $lead. If the alignment is omitted, it is set to $lag. If alignment is $lag, each output value is formed from input values in a window extending back in time, ending with the current value. If alignment is $center, each output value is formed from input values in a window centered on the current value. If alignment is $lead, each output value is formed from input values in a window extending forward in time, beginning with the current value. The threshold is useful in special situations where you may not want to use a value based on fewer points than the complete window size; if the current window has fewer valid input points than the specified threshold, the transform value is still calculated but the point's status is set to Cut. The default value of the threshold is the same as the window size.
IMPORTANT:
For alignment=$center or alignment=$lead, moving window transforms use future data values. Be careful that your model does not rely on future values that will not be available when the model is used online with current, rather than historic, data. Compensate by time-lagging future-windowed inputs or use $lag. Also note that $moveexp and $movegauss have non-standard behaviors as noted in the explanation of those functions.
NOTE:
Moving window transforms that have an unusual behavior if non-averaged.
Provide Feedback