Marketing Mix Modeling for Advertising Budget Allocation and Demand Planning
A marketer has many tools to help allocate the advertising budget across media channels. In 2020, the revenue of advertising market amounted to $ 508 billion. From experience of many companies, part of the budget is allocated based on experience, which may not be very efficient. To prevent this from happening, they use marketing mix modeling. But back in 2016, they wrote on the Internet that this tool is dead.
Together with Alexandra Goncharova, a Data Science expert from IBA Group (Minsk), we’ll figure out what marketing mix modeling is accused of and whether it has a right to exist.
Table of contents
What is Marketing Mix Modeling?
Marketing mix modeling (media mix modeling, or MMM) are mathematical models that help measure the impact of past and predict the impact of future marketing activities on sales and profits.
Marketing mix modeling uses aggregated data such as sales or marketing budgets over several years to evaluate the efficiency of conventional and digital promotion channels. In addition, with the help of marketing mix modeling, marketers can also take into account external impact factors: seasonality, trends, competitors’ actions, etc.
Marketing mix modeling became trending back in the 1960s and 70s, when the market was quite simple: there were several promotion channels, there was no accumulated information, and technologies just started to develop. For example, when Kraft launched Jell-O gelatin desserts, they have to choose between three or four television channels and magazine advertisements to promote. Currently, companies are considering dozens of channels on TV, radio stations, print and outdoor advertising, online advertising, blogs, PR, sponsorship, etc. for advertising their goods.
What Marketing Mix Modeling is Criticized for?
Today the market has become highly competitive, there is a lot of information and customers almost do not reply to advertising “for everyone”. To reach out to a customer, you need detailed and comprehensive analytics. For this reason, many believe that MMM underperforms.
Three main complaints about MMM:
- Marketing mix modeling incorporates only the short-term effect of advertising.
- Marketing mix modeling does not evaluate the meaning of the advertising message or its creatives.
- Marketing mix modeling is often not able to measure the effect of launching multiple media channels at the same time.
Why Marketing Mix Modeling Still Works?
Despite criticism, MMM works. There are several marketing mix modeling solutions on the market, for example, from Nielsen or Coffee-analytics.
In order for the marketing mix modeling to correctly answer the questions posed, two things are necessary:
- Historical data. The more information is available for analysis, the more accurate result MMM will be able to generate, which will take into account both external factors and the long-term effect of activities. For example, if the MMM model did not have data on outdoor advertising, it will not be able to predict the flow of customers when spending on outdoor advertising increases.
- Correctly selected models. Data Science offers a variety of mathematical models such as neural networks, boosting algorithms or tree-based algorithms, etc. Thus, in order to optimally distribute the budget across media channels, determine their factors and the effect of overlap, one should use a model that quickly integrates into the tools, provides good performance when there is not a lot of data, and where the results are easy to interpret.
baselinet — assessment of the base value of KPI, subject to slight fluctuations in media activity;
Mediai,t ~ media activity (audience, budget);
βi.γi — response function parameters;
есоnj,t — economical forces;
αj — parameters of economic forces;
εt — model intercept;
AdStock (xt. λi,) — adstock function that simulates autoregressive processes (processes that include the “decay” effect), where λi is the decay parameter
Model covering different media channels and additional factors
The creative component of the ad is assessed in other ways, such as A/B testing.
Alexandra Goncharova shared several MMM projects that she and her team have completed over the past few years.
Example of Marketing Mix Modeling for a marketing company: “halo effect” and a calculator of optimal budget allocation
The company wanted to know whether the budget for drug advertising is being spent correctly and to prepare an efficient advertising strategy.
Alexandra’s Data Science team, together with the customer’s professionals, analyzed advertising costs in different channels: mass media, TV and radio, outdoor advertising and online campaigns. They had data for five years. This allowed them to introduce economic factors and seasonality in the forecast model.
The model showed that outdoor advertising does not bring the desired result and needs to be treated more attentively.
Advertising for one drug does not increase the sales of other drugs from that manufacturer. The absence of such an effect, also called the “halo effect”, is understandable in this situation — we often remember the name of the drugs, and not the manufacturer.
Marketers also received a dashboard that assists them in comparing sales and ads across different channels, as well as a calculator for optimal budget allocation for the desired period. This helped to determine the limits above which it is pointless to invest in advertising: new customers will not bring profit.
Marketing Mix Modeling for the Lidskoe Pivo company: distribution volume and optimal discounts
Lidskoe Pivo wanted to analyze the quality of machine learning in the sales forecasting problem. For this purpose, data was selected containing information about permanent and promotional discounts without taking into account marketing activities. From all the information provided, data was selected for a group of products to assess the cannibalization effect. IBA Group has implemented a pilot project using marketing mix modeling and Data Science. The project was implemented in several stages.
- Data preparation. The customer provided data indicating the depth of permanent and promotional discounts, structured by partners and SKUs, permanent and promotional sales. This data resulted in a usable format. Then they collected and analyzed additional factors.
- Building an econometric model. Experts made a breakdown of indicators, determining the seasonality and the impact of additional factors, assessed the lagged effect and decay effect for discounts, as well as the effect of cannibalization within a group of goods.
- Building a forecast model that estimates the volume of distribution based on the provided amount of data.
As a result, they described the factors influencing the volume of distribution, including the sale of similar goods in a different volume and packaging; built a model that shows the impact of each of these factors; assessed the forecast values of the distribution volume for the quarter for individual products and the category as a whole.
Deviation from the actual values of forecast indicators of the model in terms of distribution volume at the category level did not exceed +/- 3%. And for individual products, the resulting value was twice the value that the customer expected to see.
In this problem, neural networks are just a tool that helps people better understand data and make the right decisions. For it to work it needs to be adjusted to the company’s business processes; we need to collect and structure the required data, and understand what factors need to be considered. You can’t do without a human here.
In addition, neural networks do not put forward hypotheses for verification; they will not be able to accurately determine the exogenous factors that need to be taken into account. Fundamental decisions should be made by a professional, and machine learning can be a convenient tool to do this.
Therefore, the main decisions should be made by a professional who will assess the adequacy of the results obtained by neural network, taking into account the current situation.
Example of Marketing Mix Modeling for a telecom operator: optimal budget allocation and factors of media channels
The telecom operator’s marketers needed to attract the maximum number of people to the points of sale. The budget for the promotion was determined in advance, and it was to be properly distributed among the channels. The company has data for the last three years: advertising budget by week, channel summaries and the number of acquired customers.
To address this business challenge, a model was developed that accommodates the basic values of KPIs, media activity and economic factors, if any. The impact coefficients of each media channel, the value of fixed factor, the coefficients of lag and decay were determined, and the impact of media channels was assessed.
As a result, the new approach attracted three times more customers.
It helped determine the factors of each media channel and its efficiency. For example, online advertising is efficient, hits without delay (lagging) and without decay. At the same time, advertising on TV is efficient with a long decay period and an initial delay of impact.
We developed an optimal budget calculator. You just need to set the required amount, and the calculator will suggest how much to invest in online advertising or TV advertising. You can also now estimate ROAS (return on ad spend) — how much income each invested dollar brings.
From our own experience, we have proven that marketing mix modeling has a right to exist. It will take several months to develop a model, but you will be able to re-evaluate the performance of an advertising campaign, calculate the return per dollar invested and increase ROMI.
Marketing mix modeling turn accumulated data into an asset that helps you outperform the competition.