Bayesian Methodology
Bayesian MMM at Rockerbox
What is Bayesian MMM?
Bayesian MMM applies Bayesian statistical methods to marketing mix modeling. Unlike traditional statistics that view probability as frequency, Bayesian statistics treats probability as a measure of belief or evidence, incorporating both prior knowledge and current data.
Why Bayesian for MMM?
Bayesian methods are ideal for MMM because they allow the inclusion of prior industry knowledge, adapt to various data types, provide uncertainty estimates, and handle complex models effectively. Though computationally intensive, the insights gained are invaluable for nuanced marketing data analysis.
Methodology Insights
Handling the Halo Impact
Bayesian models account for the halo effect, such as the impact of TV spend on Search, by incorporating prior knowledge, modeling cross-channel interactions, considering time lags, and using hierarchical models. Regular updates ensure the model adapts to new data, maintaining relevance and accuracy.
Human and Machine Roles in Model Building
Our MMM process is largely automated, requiring initial configurations before triggering model training. Human involvement is focused on ensuring model fit and realism. We iterate on the models to best represent the media mix and KPI relationships, enhancing them with machine learning techniques.
Seasonality in Models
We account for both regular and holiday-related seasonality in our models. Our future plans include integrating tools like Prophet to capture these seasonal effects more comprehensively, offering more detailed insights.
Evaluating Model Effectiveness
We use model fit metrics like R^2, MAPE, and confidence intervals to evaluate our models. These measures help us understand the accuracy and reliability of our predictions, ensuring that our models provide realistic and actionable insights.
Model Creation and Testing
Our models are tailored to each client's unique data and needs. We validate models through testing against actual outcomes, adjusting the test_start_date to ensure true predictive performance. This approach ensures our models are both accurate and applicable.
New Channel Testing
When integrating new channels, we assess their fit and impact on existing models. This process involves running the model with the new channel included and observing the effects on overall model performance and accuracy.