EvaluationWindows

class pymc_marketing.mmm.counterfactual.EvaluationWindows(windows, max_window, dates)

Per-period evaluation windows, and the cutting of arrays to fit them.

A counterfactual is evaluated on a window around each period rather than on the whole date axis, because the alternative is re-evaluating the entire series once per (period, channel) pair. The window reaches back l_max for carry-in history and forward l_max for carry-out, and is clamped to the fitted dates rather than padded out to those ideal bounds. Clamping is what makes a windowed evaluation agree with a full-axis one on every evaluated date: at the start of the data there is no history to supply, the graph’s own adstock already pads with zeros there, and padding instead would inject synthetic history – inert for spend, but not for an effect whose contribution at zero spend is its own intercept.

The dates that are summed follow the adstock’s convolution mode. Trailing decay is attributed from the period forward, leading effects from the period backward, and overlap effects to both sides. Under full-axis evaluation, After deliberately still excludes moves before the period: those can be caused by a date-reducing node rather than adstock kernel mass, and including them would make periods overlap in a quantity callers read as a decomposition. For Before and Overlap, however, dates before the period contain real leading kernel mass and are included in the sum.

Parameters:

windowslist of PeriodWindow

One per period, in period order.

max_windowint

Longest window. Shorter ones are right-padded to it so they stack for batched evaluation; a causal filter cannot reach back into that padding from any evaluated date.

datespd.DatetimeIndex

The full fitted date axis the windows index into.

Methods

EvaluationWindows.__init__(windows, ...)
EvaluationWindows.build(*, periods, dates, ...)	Work out the window of each period, and which of its dates are summed.
EvaluationWindows.build_scenarios(*, ...)	Build the perturbed spend arrays that have to be evaluated.
EvaluationWindows.cut(values, dtype)	Cut a date-first array into one padded window per period.
EvaluationWindows.time_index(dtype)	Build the positions along the fitted axis that each window covers.

Attributes

EvaluationWindows.windows
EvaluationWindows.max_window
EvaluationWindows.dates