Optimizing brand campaign spends using machine learning (ML)

When a company starts TV advertising, there are two important parameters used to determine the effectiveness of the campaign. Direct response and Brand response.

Direct response is a proxy for ‘persuasion’.

Brand response is a proxy for ‘brand appeal’.

While the Direct response is still easier to measure (comparatively), measuring brand response is where I have seen many advertisers struggle.

But first, why is it important to measure brand response?

Given the limited marketing budget with startups and many initial stage companies, they cannot rightfully think of doing continuous TV branding.

Hence, they have to go with ‘Flighting’ or ‘Pulsing’ schedule of running TVC.

Each method has its own pros and cons.

If a company decides to go for the Flighting technique, the major con is the likelihood of wear out and hence you have to decide when to resume your TV campaigns again to build brand appeal before complete wear out.

Hence, it is very important to measure Brand response and then be able to predict ad stock decay rate; simply put, when their ‘brand appeal’ will wear off.

Below is a typical graph of a ‘successful’ TVC campaign where we see a baseline shift in our metrics pre and post-campaign.

So, what are we really trying to solve?

1. Determine when the ‘brand appeal’ of a campaign will get over
2. The right time to start advertising again so as to make the best use of our marketing budgets

Talking to most agencies and companies, we realized that industry-standard ad stock decay rate is 2 weeks, which was not the correct way, as decay rate is a function of campaign messaging, how the creative performed, media planning and 10 other factors come into play.

So, what could be the right way? This looked like a perfect example of supervised learning where we can solve the problem using regression analysis.

Then I stumbled upon an article by analyticsartist that explained the mathematical way of determining the ad stock decay rate.

For simplicity sake, we used single variable linear regression analysis which can be further stretched out to multivariate analysis, the single variable being ad spend (a proxy for ad exposure)

Here is the output -

Applying it to our company’s data — we were able to predict that the impact of the TV campaign and that ‘brand appeal’ will wear off by the end of week 6 (as opposed to 2 weeks) and predict the right time to resume TV campaign.

You can reach out to me for the python code and use the formula from the above link, to measure ad stock exposure and predict ad stock decay rate to time your TV campaigns and optimize your marketing budget.

Happy optimization!