Bayesian Media Mix Modeling as a marketing strategy tool

In the product marketing process, which we usually call marketing, making decisions such as: determining targets or goals, choosing the use of marketing channels or channels, the length of time for marketing, and the budget or capacity for costs to be incurred are essential things to plan appropriately. Creating a marketing strategy based on intuition alone is not a wise thing to do. But making it only with simple data analysis calculations is just as bad, especially when it involves significant marketing funds or budgets.

Often a company's fundamental problem in the marketing process is finding the best way to allocate marketing budgets across various media channels. How should marketing funds be allocated to radio, TV, email marketing, social media like Instagram and Facebook, video ads like YouTube, search engine ads like Google, or even newspapers?

One way that is often used is heuristics, which is to determine the most reasonable (according to personal opinion) rule of thumb about what might be most appropriate for your business. It simply sets a marketing budget based on a percentage value of expected revenue. In the hotel sector, this practice is still often used. When determining the marketing budget, they will allocate a certain percentage from the predicted revenue target. But this model involves guesswork or assumptions, which is better avoided in doing good planning, regardless of the size of the marketing budget.

The solution for determining an accurate company marketing strategy is a Bayesian approach or modeling in the practice of Media Mix Modeling (MMM).

In the world of marketing analysis, the term Media Mix Modeling (MMM) has long been known. Quoted from the journal "Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects" (Jin et al., 2017), MMM is a statistical analysis such as multivariate regression on sales data and marketing time series to estimate the impact of various marketing tactics (marketing mix) on sales and then forecast its future implications. This calculation is often used to optimize different types of advertising and promotional tactics concerning sales revenue or profits.

This technique was developed by econometrics and was first applied to consumer packaged goods, as the producers of those goods have access to more accurate data. Then, the increase in the availability of data (big data), computing power and technology that is now much more sophisticated, and the pressure to measure and optimize marketing spending has finally driven the explosion in popularity of using MMM as a marketing tool.

Advertisers use MMM to measure advertising effectiveness and provide an overview of future budget allocation decisions. However, advertisements usually have specific characteristics that are a bit inconvenient, such as results that will decrease due to the influence of time constraints and the different aspects of each marketing channel, which are difficult to capture using ordinary linear regression. The Bayesian approach is also used in the MMM technique to overcome problems that cannot be measured by the usual method.

Bayesian statistics are not a new phenomenon. Its name is taken from the originator of the theorem, namely Thomas Bayes. This approach was first published in the 18th century. Unfortunately, technological developments did not support the implementation of the Bayesian method at that time. So this approach was ignored. Only in 2017 did this statistical approach begin to be discussed again after being promoted by Google.

What made statisticians dislike Bayesian methods back then was the number of calculations they needed to complete, making frequentist interpretation a much more practical option.

Today, the computing power of computers combined with powerful algorithms, such as the Markov Monte Carlo chain, makes it possible to harness the potential of Thomas Bayes' ideas.

The Bayesian Modeling formula is as follows:

P (A | B) = { P (B | A) . P (A) } / P (B)

Where P (A | B) is the probability of A occurring given that B has occurred and vice versa for P (B | A). In contrast, P (A) and P (B) are the probability of A occurring and B occurring.

We can understand Bayesian MMM as regression modeling applied to business data in simple terms. The aim is to estimate the impact of marketing activities and other factors that may influence them, such as the number of new customers over some time.

For example, in Bayesian MMM, to be able to estimate the impact of marketing activity, two main predictor variables can be used, namely:

1. The spending rate for each media channel over time.
2. A set of controls in specific criteria can capture differences in values ​​due to seasons or economic indicators etc.

Rather than modeling the number of customers acquired as a linear function of marketing expenditure, Bayesian goes a step further by modeling the saturation potential of different channels. For example, in a simple marketing budget modeling, 5% of revenue is targeted to get 1000 new customers, and this target is usually always the same, rarely changes, and ignores other factors. At the same time, advertising channels or channels have different characteristics that need to be considered before making budget and target decisions. Some channels have different audience criteria in different seasons, so the target for each channel is distant. The results at certain times of promotion may be different from the results of promotions at other times, and the second promotion will have different results from the first promotion on the same channel.

When thinking about optimizing a channel, considering the saturation level of the audience is very important. That way, advertisers know when to quit and significantly improve customer acquisition and spending for that channel. Knowing the saturation of each channel is very important in making future marketing spending decisions; Bayesian MMM can include this criterion in its statistics to obtain a more accurate predictive picture.

It is essential to understand that marketing for a particular channel may have short-term or long-term effects. Do you remember the advertising jingle on TV that was there 1-20 years ago? Those would be examples of significant long-term impact if viewers placed an ad several years ago. Taking time control into account is essential for showing that a channel has an effect that turns out to be only "short term" to plan to do more marketing on that channel. On the other hand, advertising within a larger budget but only in one or two rounds may be the right solution for tracks that have a long-term effect.

An illustration of the influence of reach, budget, time, and control criteria data in marketing is as illustrated in the following graph:

With the help of this approach, the complex marketing process can be calculated precisely in a program that was deliberately created using the Bayesian MMM formula.

The data will be entered into a program that uses Bayesian MMM calculations. The program then generates a picture of the information. This description of data or insights can be validated and refined on an ongoing basis with hands-on tests. Validation and refinement tests will eventually produce new data.

The advantages of using Bayesian MMM can be described into three significant benefits as follows:

1. They are adding elements of human knowledge in the form of old knowledge data. Prior knowledge about the effectiveness of marketing channels can come from various sources. One of the data sources could be the knowledge of a marketing manager who has accumulated hundreds of marketing campaigns, their understanding of changes in the prices of competitors' products, etc. When the marketing manager of a company stops working, the knowledge data he has does not necessarily stop and disappear because it has been collected and inputted into the program. Another source of prior knowledge could be the result of carefully conducted field experiments such as product tests, surveys, sales results, etc. The Bayesian approach allows prior knowledge to be elegantly incorporated into the model and measured by an appropriate mathematical distribution.
2. They are optimizing the merging of old knowledge with new data. The essence of the Bayesian approach is keeping our understanding of the data up to date. Models will no longer rule out human knowledge; they add to the Bayesian method.
3. It increases the significance of the data. Another benefit of the Bayesian approach is that it reduces misinterpretation of data and information. Indeed, companies need to make data-driven decisions. However, this needs to be done with existing data records to provide an accurate picture. The more control parameter data entered, and the more knowledge added, the more precise the information description and predictions would produce.

So has your company used a Bayesian Media Mix Modeling-based marketing program or tool?