Implementace a porovnání modelů marketingového mixu v knihovnách Robyn a Meridian

This thesis examines the implementation and comparison of two approaches to Marketing Mix Modeling (MMM) using the open-source libraries Robyn and Meridian. Robyn, developed by Meta Platforms in the R programming language, is based on a frequentist framework using ridge regularization and optimization through the Nevergrad evolutionary algorithm. Meridian, developed by Google in Python, implements a Bayesian approach utilizing prior parameter distributions and posterior sampling via Markov Chain Monte Carlo (MCMC). Both models were applied to an identical dataset capturing daily marketing expenditures and conversions of an anonymous company operating in the United States restaurant sector during the period 2021–2024. The aim of the thesis is to analyze the differences between the two approaches in terms of statistical methodology, predictive accuracy, interpretation of channel contributions, and practical applicability of the results. The findings indicate that the Meridian model achieves higher predictive accuracy on the holdout sample compared to the Robyn model, while both models consistently identify Meta Facebook as the channel with the highest absolute contribution to purchases and Google Shopping, TikTok, and Meta Instagram as the channels with the highest ROI. The Bayesian framework of Meridian provides more explicit uncertainty quantification through credible intervals, although the results remain sensitive to prior specification. Robyn’s frequentist approach offers faster implementation and more straightforward scenario testing, but does not provide direct probabilistic interpretation of parameters. The thesis contributes to the methodological discussion on selecting an appropriate MMM approach depending on data availability, the analytical capabilities of an organization, and the nature of the decision-making problem.