mmm budget optimizer

Marketing Mix Modeling (MMM) projects often fail because they rely on expensive, black-box software or noisy Bayesian estimates that are hard to verify. This skill provides a deterministic C++ engine to attribute sales to specific marketing channels and reallocate budgets based on marginal ROI. It handles the full pipeline from data preparation to optimization with cold, mathematical precision.

What it does

• Data sanitization processes raw sales and spend files into cleaned weekly CSVs while checking for data gaps or singular matrices.
• Model fitting uses a C++23 engine to fit adstock and saturation curves, measuring accuracy against a 13-week holdout set.
• Attribution reporting calculates both total and marginal ROI, identifying which channels have no measurable effect on sales.
• Budget optimization reallocates weekly spend across channels to maximize sales within specific safety bounds.
• Performance forecasting predicts future sales lift based on the optimized spend plan.

How it works

1. Prepare by merging raw CSVs into a standardized weekly format using the internal Python scripts.
2. Fit the model using the mmm CLI to generate a model.json and verify Holdout MAPE and R² scores.
3. Optimize by running the engine to find the highest-ROI allocation for a given budget and bound.
4. Report by generating a Markdown summary that lead with marginal ROI and attributes sales to specific channels.

Frameworks & tools

Requires a C++23 compiler (GCC 13+ or Clang 17+), CMake 3.27+, and Python 3.9+. It utilizes a dedicated C++ engine for deterministic fitting without relying on external API keys or cloud services.

Why this beats prompting it yourself

Generic LLM prompting cannot perform the complex non-linear regressions required for adstock and saturation modeling. This skill enforces strict validation gates, ensuring you never make budget decisions based on models with poor holdout accuracy or insufficient training data.

Use cases

• Budget reallocation to move money from low-marginal-ROI channels to high-performers.
• Channel auditing to see which marketing activities are actually driving sales versus baseline demand.
• Media planning to forecast the weekly sales lift of a 50% increase in total marketing budget.
• Incrementality verification to prepare for geo-lift or holdout tests based on model attributions.

Known limitations

Requires at least 104 weeks of data for reliable results. High-confidence optimization is limited to +/- 50% of current spend to avoid unreliable extrapolation.