Marketing Mix Modelling (MMM) or Media Mix Modelling is an analytical approach which uses statistical techniques to quantify the impact of various marketing tactics (Market Mix) to optimise and forecast promotional tactics with respect to sales revenue or profit.

MMM helps figure out an optimal spend allocation. It is the process of quantifying the impact of each marketing vehicle in terms of ROI and effectiveness.

In MMM, the volume of sales is modelled as the dependant variable, while the independent variables represent elements (drives marketing) of the marketing process.

The objective is to evaluate past marketing performance ( 2-5 years of historical data, monthly or weekly ) to measure how media channels drive sales to determine true marketing ROI and optimal mix of media activity.

Trend Analysis: before and after

We look at the difference of revenue between before a marketing event (like newspaper ad) and after the marketing event to determine if it drove a change in revenue, and analyse what that change is.

Before and after analysis is the first and best method to try when you are attributing activities.

Linear regression with single variable

In order to know how a variable affects sales, linear regression is a good starting point. You can build a regression model directly into Excel or LibreOffice. In Excel, you can use the LINEST function (linear estimator).

Variables with positive and negative correlations

Correlation isn't causation. However if a variable is not correlated to another variable, it probably won't affect much the second either.

To calculate correlation between two columns in Excel, we can use the CORREL function.

To show how many variables are correlated to each other, we can create a correlation matrix. Remember how correlation between r.v.s $X$ and $Y$ is defined:

so, if we have a vertical vector $X=(X_1,\dots ,X_n{)}^t$, then the covariance matrix is:

and the correlation matrix follows naturally.

Multivariate Regression

In order to accurately attribute the impact of each channel on sales, you need to account for multiple variables and their interaction with each other.

LINEST function in Excel can do that too.

To calculate the contribution to the revenue of the media channels after we have built our multivariate regression model, we just multiply the respective column (e.g.: social ads) by the respective coefficient determined by the linear regression; that is the contribution to the revenue of that particular channel. It is basically how much the channel made in revenue according to our model. The percentage is of course that, divided by the total predicted revenue, which is just the sum of all the contributions, plus the baseline (the intercept).

It's important to visualise the error of the model, in order to determine if there is some pattern: that may mean that we are missing something important that the model doesn't capture.

Diminishing Return and Adstock {#Diminishing Return and Adstock}

Not all variables have a linear relationship with sales: for example, brand marketing tends to have a lagged impact long past when the ad ran.

Also, what we often see is a diminishing return effect on the incremental dollar spent