Media Mix Modeling to the Rescue

Kevin Hillstrom has challenged his readers to contribute suggested solutions to the problem: how can multi-channel marketers assign sales to the different channels that they manage?
He also makes a file with 10,000 data points available to the readers and some folks seem to have taken the challenge as far as actually starting to develop some analytical models. I’ll take a simpler and more prescriptive approach here: curious as I am to read what suggested solutions readers may come up with, I am inclined to recommend an econometric model.

These models (also known as "Media Mix Models" or, more generally, as "Marketing Mix Models") resort to time series analysis to establish a causal relationship between the independent and the dependent variable. In this case, sales form the dependent variable and the budget allocated to each channel would correspond to each independent variable. The analytical method produces an expression that allows the modeler to evaluate the contribution of each channel to the outcome over a period of time.

Although the models are developed by analyzing historical data, they can forecast sales results and channel contributions given a budget or marketing plan.

Media Mix Models involve a very specific and high-value skill set to develop, and although well documented in Academia, require great care in development if they are to have some adequate business accuracy.
Aside from being able to determine the seasonality and other cyclical behavior inherent to the business outcome (sales, web visits, calls to a queue, etc.), when properly developed these models also account for:

• Lags between stimulation (marketing campaign touching a prospect) and response (sale)
• Synergy between media; for example, TV and DM produce higher sales when executed  together and with a given lag than the sum of the results obtained from each one being executed separately
• The impact of macroeconomic and other extraneous factors on the business outcome; for example, some non-essential items may be sensitive to unemployment rate (or only in some seasonal periods)

Regarding Kevin’s problem, I’d say that the variables that he presents would have to be evaluated for their impact on the outcome as well as for multicolinearity, the resulting model should clearly indicate the contribution of each factor to sales.

Marketing Mix Modeling is at the heart of the more general Marketing Mix Optimization and is important tool in Accountable Marketing

Comments

[li wang]

it's a data mining problem, I'm familiar with using bayesian network to solve this probalem, Bayesian network is a probability model, for example, assuming the customer has rcieved e-mail, postcard and catalog, last purchase is 100 pounds,... by learning historical records, the model will compute the probability that the next purchase over 100 pounds is 80%....