A three-year weekly dataset, five media channels, one Ridge and PyMC-Marketing model with adstock decay and Hill saturation curves, checked by holdout, Monte Carlo and time-series cross-validation before it's allowed to say where budget should move.
Which channel actually earned its budget?
TV took the largest share of a three-year, five-channel spend, but a raw correlation between spend and sales can't tell a channel's own effect apart from a trend or a competitor moving at the same time.
Model each channel's real, saturating effect.
Adstock decay carries a channel's spend forward across the weeks after it runs, and Hill saturation curves cap its return, so the fit separates each channel's own diminishing effect from everything happening around it.
A 5-fold time-series split scores every fold forward in time only, never validating a week against data from after it.
The results are real; the money behind them isn't. Every dollar comes from a fixed random seed, not an advertiser's ledger.
generate_data.py builds all 156 weeks with a seeded random generator, so TV's -11.9% ROI and email's +1,152% aren't a real company's numbers, they're what the model recovers when the true relationship is already known. The question is whether the method finds that answer, not the dollar figures.