Marketing Mix Modelling Project

Project Overview

A Marketing Mix Models (MMM) has become a crucial tool for understanding the incremental impact of marketing channels on business outcomes, and identifying the optimal allocation of media budgets. In the coming years, consumer privacy laws such as the California Consumer Privacy Act (CCPA) and the General Data Protection Regulation (GDPR) are expected to expand across North America, which will limit the availability of data and degrade the effectiveness of performance-based measurement methodologies. MMMs enable business leaders to preserve user privacy while measuring marketing impact, optimizing budgets, and connecting marketing spend to business outcomes, even without cookies.

For this analysis, I implemented a Bayesian Regression approach using Google’s LightweightMMM framework [1], leveraging the comprehensive API documentation provided in the repository [2]. The model provides a scalable and interpretable solution for measuring media effectiveness and optimizing marketing investments. Readers can reference my Github Repository, for the complete implementation documentation.

This analysis represents a foundational implementation of Google’s LightweightMMM framework using simulated data. While the underlying Bayesian approach remains robust, readers should note that this model reflects techniques and data structures from 2022 and serves as a methodological reference point. For production implementations, consider utilizing the latest models and methodologies.

Executive Summary

Optimization Results: 6.3% Brand Sales Improvement Through Strategic Budget Reallocation

The media mix optimization simulation reveals a meaningful opportunity to improve performance by reallocating budget across channels. The model simulation predicts a 6.28% increase in target variable through strategic budget reallocation.

MMM-Optimal-Budget-Simulation

Budget Reallocation Recommendations

Increase Investment:

  •  TV: +5% budget increase (maintaining dominance as top performer)
  • Direct Mail & Digital Out of Home: +1% budget boost to capitalize on proven mid-tier performance

Decrease Investment:

  • Online Video: -3% reduction (17% → 14%) despite strong historical performance
  • Social Media: -2% reduction, suggesting saturation at current spend levels
  • SEM, Billboard, Online Display: Minor adjustments reflecting optimization at the margins.

Business Impact

If the channel budgets were optimally allocated, it would result in ~1.5 million more brand sales over two years. The company would have seen their target KIP increase from the reported value of 22,358,139 to 23,854,663.

The 6.3% performance lift represents meaningful incremental value with minimal risk, as all changes stay within ±20% bounds of historical spending patterns. This conservative optimization approach ensures implementable recommendations while capturing measurable improvement opportunities across the media portfolio.

Unknown

Strong Brand Equity Foundation (blue): The substantial blue baseline contribution demonstrates robust organic demand, indicating strong existing brand equity that provides a solid foundation for paid media amplification.

TV’s Sharp Impact (gold): The sharp spikes in TV contribution clearly correspond to campaign flights, showing TV’s ability to create immediate, measurable lift above baseline. These spikes demonstrate TV’s power for driving short-term activation while the broader seasonal curve shows sustained impact.

Model Fit

These outcomes of the model fit suggest that the model is performing well overall, with strong predictive power and relatively low error. 

Unknown-1

  • An R-squared of 0.902 indicates a strong fit, suggesting that the model is able to predict the KPI very accurately, leaving only a small portion of the variability unexplained.
  • While the model’s predictions are generally accurate, there is still an average error of roughly 8.731% between the predicted and true KPIs.

The strong fit could represent genuine model quality, however I think this is unlikely given this dataset likely has more predictable patterns than real-world data.

Dataset Details

This model utilizes a dataset generated with the Aggregate Marketing System Simulator (AMASS), an open-source R package provided by Google for research purposes [3]. The AMASS tool simulates realistic marketing datasets. The source code for the AMASS package was obtained from the Google GitHub repository.

  • Initial Data: AMASS generated 228 rows of weekly data, representing 4 years.
    Normalization Period: The first 104 rows were removed to allow the simulator time to normalize.
  • Final Dataset: The resulting dataset contains 104 rows, representing 2 years of weekly data.

This final dataset captures a balanced media budget allocation across traditional media, social media, search engine marketing (SEM), programmatic display advertising, direct mail, digital out-of-home (DOOH), and billboards.

Target Variable:

This metric tracks brand sales volume as the primary key performance indicator for measuring business success.

Media Variables:

The dataset includes the following media spend variables:

  • TV Spend
  • Online Search Spend
  • Programmatic Spend
  • Social Spend
  • Connected-TV Spend
  • Direct Mail Inserts
  •  Digital-Out-Of-Home Spend
  • Billboard Spend

Additional Variables:

Merge weekly records with Canadian national holidays across the 52 weeks of the year, and seasonal variables to indicate seasonal consumer behaviours. 

References

[1] Duque, P., Nachbar, D., Abe, Y., Ahlheim, C., Anderson, M., Sun, Y., Goldstein, O. and Eck, T. (2022) LightweightMMM: Lightweight (Bayesian) Marketing Mix Modeling. https://github.com/google/lightweight_mmm

[2] Duque, P., Nachbar, D., Abe, Y., Ahlheim, C., Anderson, M., Sun, Y., Goldstein, O. and Eck, T. (2022) LightweightMMM Models API Documentation. https://lightweight-mmm.readthedocs.io/en/latest/api.html

[3] Zhang, S. and Vaver, J. (2017) Introduction to the Aggregate Marketing System Simulator. https://research.google/pubs/introduction-to-the-aggregate-marketing-system-simulator/