PeriodWindow

class pymc_marketing.mmm.counterfactual.PeriodWindow(start, end, in_window, actual_dates, in_eval, eval_dates)

The stretch of fitted dates one period is evaluated over.

Two nested date ranges, and the distinction between them is load-bearing. The window is what the graph is handed, wide enough on both sides that the perturbation’s whole effect is computed correctly. The evaluation dates are the subset whose differences are summed into the increment. Both are recorded here, in the two forms the rest of the code needs them in – a mask over the full date axis, for slicing a full-axis baseline, and the dates themselves, for labelling coordinates – so that no caller has to recompute either from l_max and a frequency offset and hope it lands on the same answer.

Parameters:

start, endpd.Timestamp

Bounds of the period itself – the dates the counterfactual factor is applied to, as opposed to the wider window it is evaluated over.

in_windownp.ndarray

Boolean mask over the full date axis selecting the window’s dates.

actual_datespd.DatetimeIndex

Those dates, in order.

in_evalnp.ndarray

Boolean mask over the full date axis selecting the dates that enter the sum. A subset of in_window, by construction.

eval_datespd.DatetimeIndex

Those dates, in order.

Methods

PeriodWindow.__init__(start, end, in_window, ...)
PeriodWindow.build(*, start, end, dates, ...)	Derive both date ranges from the window mask and the carry-out end.
PeriodWindow.eval_mask(max_window)	Positions in the padded window that enter the sum.
PeriodWindow.offsets_within(first, last)	Positions inside the window of the dates in [first, last].

Attributes

n_actual	How many fitted dates fall in the window.
start
end
in_window
actual_dates
in_eval
eval_dates