The Use of Marketing Models and Evolutionary Algorithms to Optimize Advertising Campaigns

Based on the paper “Differential Evolution Framework for Budget Optimization in Marketing Models with Saturation and Adstock Effects” by Morais et al. (2024), available at ScienceDirect.

Marketing models are statistical and mathematical methods that aim to elucidate the true results of advertising, and even though these models are not as popular as Large Language Models (LLMs) and the new Generative AI models, they are just as important. We can, for instance, use a multi-touch attribution model to estimate which media channel is critical to conversions or evaluate which advertising piece performs the best with a controlled A/B test. But, when it comes to understanding and optimizing the media's impact, the need for a Marketing Mix Modeling is certain. Building a model such as this seems easy at first glance, since there are a lot of open-source tools out there to help with that task, but this is where many companies often make mistakes: creating those models without proper understanding of the concepts involved will most probably lead to misinterpretation and consequently to a wrong advertising strategy.

This type of modeling, commonly called MMM, is among the most important marketing tools and is usually built upon multiple linear regressions applied to time-series forecasting models while considering advertising carryover and saturation effects, concepts very important on the understanding of how the advertising works. This first effect, also called "advertising adstock", is a concept studied since the 1960s that relates to the "accumulating and decaying" aspect of ads. In simple terms, every piece of advertising may carry both an immediate effect and a long-term effect that decays over time, implying that successive ad investments may create a virtual effect of "accumulated investment," which can lead to an outcome totally different from what is expected. As for the saturation effect, also a concept studied since the 1960s, it relates to the diminishing returns that overexposure to an ad may cause. Basically, every media channel has a saturation point, indicating the moment when an increase in advertising investment won't lead to an increase in the response, meaning that everything past that point is just overspend. A point often neglected is that the saturation curve is a function of the response by the virtual investment (adstock), not the real investment value. This means that successive advertising investments may cause a media channel to saturate, even if the immediate investment value is low.

As we can see, there are a lot of variables to consider in order to optimize a marketing campaign. The fact that the accumulating effect of the ads must be considered when looking at saturation means that not only the share of the budget between the channels is important, but also how each channel's budget is allocated over time, since the same investment value may lead to different outcomes depending on the distribution of the insertions. The sum of these factors makes this a problem with multiple solutions. And how can we calculate what's the best distribution for the budget considering all of this? Through an optimization algorithm, of course! These are algorithms that aim at finding the optimal solution for any given problem, and have many categories and applications. There are, for instance, swarm algorithms, which are based on populations and usually mimic the behavior of animals in nature, and there are also evolutionary algorithms, which are based on the process of biological evolution and natural selection and uses the concepts of evolution through generations and survival of the fittest to reach the optimum solution. The paper on which this article is based proposes a solution that uses the Differential Evolution algorithm, an evolutionary algorithm derived from the classic Genetic Algorithm known for its versatility and ability to avoid local minima.

The process of optimization aims to find the best value to be invested in each media channel and also how this value should be distributed over time. To achieve this, we must chain two simultaneous differential evolution algorithms: the first algorithm performs the “global optimization”, receiving the budget available for the campaign and suggesting possible distributions of this value among the channels. The second algorithm, then, performs the “local optimization”, receiving the value allocated by the global one and searching for the distribution over the desired time that will grant each media channel the best Return on Ad Spend (ROAS) while calculating the adstock and saturation effects. The global algorithm then receives back the result of the local algorithm for each media channel and performs a new distribution based on the total result, and this process continues between local and global optimization until the best solution for the system is reached. This process, when correctly applied, has the capability of getting the most out of marketing campaigns, achieving the best outcome sometimes even while spending less.

This is an example of how concepts from multiple areas can be joined together to create solutions that can be applied in real-world scenarios. Given the large amount of money that is spent on advertising, it's unacceptable to still have marketing based on gut feeling. In fact, there is an increasing need for marketing models, but there must also be a need for the knowledge to create such models, to evaluate third-party companies' solutions, which sometimes rely on their clients' lack of information to deliver poor models, and even to evaluate the open-source tools, that provide this optimization with the MMM but most of them have flaws on their methodology, which was the motivation for this paper to be written. More detailed information about the mentioned topics and a practical application can be found in the paper by Morais et al. (2024).