There was a time when we considered traditional marketing practices and the successes or failures they yield as an art form. With mysterious, untraceable results, marketing efforts lacked transparency and were widely regarded as being born out of the creative talents of star marketing professionals.
As the world of digital marketing has exploded, it changed things. The rise of big data and a more scientific approach has been overwhelming. The new technologies, complex algorithms, and statistical applications often leave us scrambling to keep up.
To address the issue of how to manage incoming data and then use that information to make impactful decisions, a clear analytics strategy is necessary for all companies. Methodical, strategic planning in the form of Marketing Mix Modeling (MMM) can help you overcome this problem by finding the optimal marketing mix and proving the return on investment (ROI) that your marketing activities provide.
Marketing Mix Modeling
Marketing Mix Modeling (also known as media mix modeling) is a way to determine the combined effect of different marketing initiatives across several media channels. Combining data from various media variables and in a statistical model, quantifying its effect on a commercial KPI like sales – can help your business identify the most profitable media mix to invest in.
Like all marketing analytics methods, this technique has drawbacks and limitations. The general problem of attributing the effect to the activity that caused it is inherently complex and requires sophisticated models to disentangle the dynamics behind the decisions of the masses. Traditional marketing mix models are usually difficult and time-consuming to update, and they only operate on a channel level, making the insights very high level and infrequent.
Characteristics of traditional Marketing Mix Modeling:
- Based on linear regression
- Models every variable as independent
- Impossible to detect cause and effect
- Forces a linear interpretation of a nonlinear world
- Dependent on human interpretation
- Takes time – outdated learning
Learn how you can predict the marketing effects on sales, growth and profitability all without using any cookies as part of the dataset.
AI-based Marketing Mix Modeling
The complexity of analyzing and optimizing the effects of media investments and marketing activities has substantially risen over the past couple of decades, driven by media fragmentation. As a result, the task is no longer analyzing response curves and optimizing marginal costs. Instead, it is now a game of balancing exponentially high numbers of data points representing the interconnected reality of how media insertions work and contribute to sales, accounting for correlations of each new day of incoming data.
No human is a match for this complexity and multi-dimensionality. That is why AI-based Marketing Mix Modeling has gained momentum. With AI algorithms, marketing teams have a very solid foundation on which to base strategic decisions. In addition, applying artificial intelligence to Marketing Mix Modeling provides organizations with major benefits since the most advanced AI-based Marketing Mix Models are able to:
- Make predictions with 90%+ accuracy to improve ROI on marketing spend
- Deliver frequently updated insights allowing you to make agile data-driven decisions
- Show synergies across different media channels
- Model all the individual factors of macro and external data inputs that can impact sales
- Utilize super high granularity to increase operationalization, but avoid data protection restriction induced by CCPA, CCPD and GDPR due to methodology not relying on cookies
- Quantify brand drivers’ impact on baseline sales
Bayesian Marketing Mix Modeling
Combining the power of AI with Bayesian mathematics allows one to build probabilistic network models relating the activities of advertisers with fluctuations in sales. Such models take all relevant data into account and model any relevant interaction to understand what drives sales. In addition, the models quantify the uncertainty in their learnings which enables them to handle data from a large variety of sources with different quality levels.
Using Bayesian Inference, you can combine learnings from historical data with general knowledge obtained from industry experience and know-how from a large number of models built previously.
Benefits of Bayesian Marketing Mix Modeling
The Bayesian Inference methodology is particularly well suitable when you have limited data. The are several reasons for this:
- Bayesian Inference is well suited for building robust mechanistic models which reflect the known interactions between the data.
- In Bayesian Inference, different media variables can learn from each other. This means that variables with little or low-quality data are supported by learnings from similar data in the model.
- Bayesian models allow experts to put in prior knowledge, thus improving the quality of the model. In addition, the prior setting makes it possible to preload models with industry-specific parameters while still allowing the model to learn from the actual data
Benefits of a platform- and AI-based Marketing Mix Modeling
One of the key advantages of subscribing to a platform is that it is NOT project based. Instead, it is a software program, a “live” product that is continuously updated with new data and produces fresh insights as needed.
Due to a lack of recent data and implementation functionalities, project-based MMM fails to become an integrated part of the daily operation. If Marketing Mix Modeling does not become a part of the day-to-day work within the marketing department, it loses more than 50% of the potential value it can generate for the business. Out of sight, out of mind. By having access to a platform with “real-time” operational functionalities, the marketing department can utilize all the functionalities in the daily work when planning and optimizing campaigns.
How to capture the combined effects from multi-channel campaigns with Bayesian and AI-based models?
As the produced models are highly nonlinear with multiple synergistic effects, there is no naive way of isolating effects independently. Indeed this is a feature and not a limitation. Marketing, and the rest of the world for that matter, does not consist of a bunch of independent events. In reality, everything is interconnected and potentially dependent on each other. This is why the cause and effect must be re-cast into a simulation scenario.
The way a Bayesian Marketing Mix Model detects and quantifies causality is by comparing a small change to a variable of interest while keeping all other parts of the network fixed. This way, it can detect how a small change can propagate throughout the networks with all the synergistic and nonlinear effects.
Type of synergies accounted for:
- A display ad is twice as effective when combined with a TV campaign than without.
- A sale is attributed to paid search, but was this sale potentially initiated by a TV ad?
- In many MMM, website visits are identified as sales drivers. But what drove the visitors to the web page?
- Competitors have a negative impact, but how does their impact change when you lower your price?
How can a marketing mix model be validated?
The quality of the models can be assured at several levels:
- Bayesian Inference specific checks: The result of the model is checked to make sure the underlying algorithms performed optimally (e.g., results must have the right effective sample size, and parameter estimates must be without unwanted correlations).
- Statistical checks: A good model must fit well to the data and be able to predict well into the future. It is important to achieve this without overfitting. In all models, we use a holdout test to assess the statistical validity of the model. Additionally, R-squared and mean absolute error are used to check fits.
- Subject experts: The learnings of the model are inspected and checked by subject experts. Media learnings are compared to industry benchmarks and any deviations are investigated.
What other explanatory factors can be used to explain variation in sales?
In addition to paid media factors, modern Marketing Mix Models account for all relevant non-media factors as well, which can be divided into different categories:
- Marketing-specific factors
- Company-specific factors
- Industry-specific factors
- Systemic factors
By accounting for non-media factors, one is able to measure and optimize the effect of paid media with high accuracy and avoid underfitting. Furthermore, adding relevant non-media factors in the Bayesian Framework allows for lower uncertainty in the media learnings.
Adding relevant non-media variables gives insights into non-media sales drivers. For example, how are sales impacted by price changes, PR activities and competitor media spend? Additionally, having these variables in the model allows measuring synergistic effects between paid media and non-media factors.
What kind of external data sources are used to model?
- Exchange rates, Consumer Price Index, Interest Rate, Unemployment Rate, Consumer Confidence, Stock Index
- Max/Min Temperature, Sun Hours, Wind Speed M/S, Precipitation by multiple locations per market
- Gasoline & Diesel Prices
- Local Holidays & Observances
- Competitive Media Spend
A model for the future – transform your marketing reporting into actionable insight
Marketing Mix Modeling is here to stay – and for companies and modern marketeers that want continued access to marketing insights, it’s time to begin the adoption.
As artificial intelligence becomes more of the norm in data analytics, organizations utilize its power to transform their marketing departments and equip them with modern and advanced marketing mix models that provide actionable insights. This, in turn, opens the door for better decision making, a more holistic approach to marketing and increased business transparency.
Blackwood Seven can assist organizations in transforming their marketing department and release the power of modern marketing mix modeling.
Ultimately, the value should materialize as a significant sales uplift from paid media of around 20-40%.