Marketing Mix Modeling needs to estimate 30,000 parameters from just 1,000 observations, making standard regression impossible. Bayesian methods solve this by using priors to reduce the effective parameters to a few hundred, making estimation tractable. That's why Recast uses Bayesian instead of frequentist statistics.
Chapters
Should your MMM use Bayesian or frequentist methods? Here's the fundamental challenge. A typical MMM is going to be looking at 2 to 3 years of data, something on the order of 1,000 observations of sales. For a normal regression, no problem. But let's work out the math on the number of parameters you have to estimate. Let's say you have 30 channels, each with time-varying ROI, so we're estimating the ROI every day. That's something like 30,000 parameters that we're estimating against 1,000 observations. And it gets worse because those 1,000 observations are not independent. So you have even less data than you think you do to estimate those 30,000 parameters. This is where Bayesian methods really shine because Bayesian methods impose constraints on those parameters via the use of priors. Now, the priors that we use are very complicated. We impose certain kinds of shapes on the time-varying parameters and things like that. So the number of effective parameters is not 30,000, it's something much less on the order of 100s. That makes the problem tractable, and that is why we chose Bayesian at Recast.
