Accenture is one of the top global consulting companies, that provides a wide range of services and solutions across over 40 industries and different business functions in more than 120 countries. Accenture combines business and technology to improve their clients’ performance. In one of their articles they cover the development of Marketing Mix Modeling and Optimization through years including the latest techniques and implementation of granular data in MMM.
This article is a summary of how MM Optimization has changed over the last 40-30 years becoming more effective with every innovation, yet it’s not perfect, and has both advantages and disadvantages.
Promotion is vital for a business’s success, and MMO helps to make sure that different media channels are used in the most effective way by creating a statistical model and analyzing its results to ensure optimal budget allocation.
Here 3 following stages of development will be discussed
- Similar Linear Regression
- Mixed Model
- State Space Model
The article is presented as a case study of a toy manufacturer ToyCo with a $100M budget, that wants to know what part of their sales can be attributed to marketing and which ad channels have more impact on sales.
Similar Linear Regression
Marketers used these easily interpreted models in the 1980s and 1990s. When using regression models, we assume that there is a linear relationship between the dependent variable (sales) and independent variables (marketing activities and external factors that also have effect on the success of ad campaigns like demographic or macro-economic landscape). These external factors can be included into the models in such a way, that we can still see marketing driven sales separately, and MMM can also show how specific marketing tools impact sales. Calculations provide us information on the rate of change in sales for every variation in marketing activity.
In case of ToyCo we have the following results:
for every $1 spent on Direct Mail we get $2 incremental in revenue,
while for TV the incremental in revenue is $1.75.
The numbers show that the current marketing mix can be improved by increasing the amount of Direct Mail and reducing TV ads.
Regression models work well only when there is a linear relationship between sales and marketing activities, which is not completely true. Accenture points out several variable factors in advertising
- Adstock: advertising loses its initial effect on sales over time.
- Saturation: the amount of money you invest in a certain marketing activity is not directly proportional to the incremental revenue it will generate, because every next spent dollar generates less revenue.
- Lag: the effect of advertising is not always immediate, so it might take a few weeks before a consumer buys your product after seeing an ad.
If you apply these factors as well, it compensates some of the drawbacks of a linear model, but it does not tackle other significant issues. For example, if there are numerous ToyCo stores across the country, the company would prefer to optimize their marketing strategy for each individual store or direct marketing area (DMA), which is a common practice in the US today. Unfortunately it would be impossible with the above regression model, since it views each DMA as identical. To solve this problem we need a different modeling approach that can separately analyze each DMA.
Mixed Models allow to set your marketing strategy at a DMA level using the same linear relationship as above but introducing a random coefficient to improve it. The random coefficient is the incremental revenue at each DMA driven by a certain marketing activity, which later along with the fixed coefficient – the average incremental revenue across all DMAs – helps create more targeted models. In ToyCo case DMAs are considered, but the focus could be narrowed down even to a single store.
While regression model for ToyCo showed that Direct Mail drove $2 in revenue for every $1 investment, in the mixed model, we can see that
New York gets $2.50 for every $1 spent, while Chicago gets $1.50.
These results allow us to further optimize the marketing strategy, by increasing investments into Direct Mail in New York and reducing them in Chicago.
Even though mixed models provide satisfactory, targeted output, proper and precise estimation requires considerable amount of data. Another drawback is that results are based on data sets covering up to several years and computed as an average, while it is recent periods of time that are more relevant for companies.
For example, in the late 2000s, the global economy crashed, which caused the decline in marketing spends. At the same time, digital advertising platforms were on the rise with their lower cost and greater personalization. The quantity of advertising campaigns increased, and the intervals between them became shorter. The dramatic changes required rapid adaptation from marketers: they had to react faster to the latest trends, while also being able to estimate the efficiency of various marketing activities at different points in time.As a result classic regression approaches gave way to a new technique much more responsive to the latest trends and market activities.
State Space Modeling
Back to ToyCo, let's suppose their sales decline sharply due to a macro-economic situation, and the company decides to reduce their marketing budget from $100M to $75M. But they don't want this budget cut to be random: they plan to reduce the amount of advertising campaigns for a period of time that will least impact overall sales. To make this possible, we need a strategy that can account for both seasonal and geographic factors. That's exactly what State Space Modeling (SSM) allows you to do. Just like previous models SSM distinguishes baseline sales from those driven by marketing activities, but it also included two innovations. Firstly, it views baseline sales not as an average figure for the year, but accounts for how they fluctuate over the period. Secondly, it calculates the effectiveness of a marketing activity at different points in time, as opposed to finding one average coefficient for the whole period of time. As a result, rise in sales in December before holidays and the decline, that usually follows, will be attributed to seasonality instead of marketing activities implemented at the time, while the prior modeling approaches would wrongly attribute those changes to marketing mix.
Here is what we get if we compare all three models.
Linear regression shows that Direct Mail drove $2 in sales for every $1 invested
Mixed models show that Direct Mail drove $2.50 in NY and $1.50 in Chicago
State Space Model shows that in the final week of November we get $1.50 for every $1 invested in Direct Mail, but in the 1st week of December we get $3.50 for every $1 invested in Direct Mail.
The new approach also uses present data to predict future sales. For example, if ToyCo chooses to see their sales weekly, this technique will predict this week’s sales based on last week’s sales, in addition to that it constantly uses insights from any given week when planning marketing mix for the following week. Yet is also important to design marketing strategies for DMA level, which can be achieved by implementing a separate State Space Model for each DMA. It will definitely require much more resources, but it will ensure that all media channels are implemented to the best advantage of ToyCo.
This third model, though, is not the last stage of Marketing Mix Optimisation evolution, and it can further be enhanced.
Adoptomedia has been successfully implementing different models mentioned above as well as improved versions of them to help their clients increase ROMI by at least 10-20%. With CheckMedia Solution V2.0 media mix optimization can be easily automated that will save you both time and money.