Media Optimization

Developed media optimization strategies to maximize Return on Investment (ROI) across various advertising channels.

Algorithmic Media Mix Optimization

Executive Summary

Marketing budgets are finite; the opportunities to spend them are infinite. This project focused on the prescriptive side of marketing analytics: “Now that we have a model, how should we spend our $10M budget?” We developed a Media Optimization engine using non-linear optimization algorithms (BFGS, SLSQP) to finding the optimal allocation of budget across TV, Digital, Search, and Social channels to maximize Revenue/ROI.

Problem Statement

The client allocated budgets based on “last year’s spend plus 5%.” This ignored diminishing returns (saturation) and cross-channel synergies. They were overspending on saturated channels (where the next dollar brings $0 return) and underspending on high-potential channels.

Methodology

1. Response Curves (The Foundation)

You cannot optimize what you cannot model. We first built Market Mix Models (MMM) to generate “Response Curves” for each channel.

  • Shape: Typically S-curves (Sigmoid) or Hill functions.
  • Diminishing Returns: Captures the fact that the first $1k in ads works better than the millionth $1k.
  • Equation: Sales = β * Hill(Spend, K, S) where K is half-saturation point and S is slope.

2. The Optimization Problem

  • Objective Function: Maximize Sum(Sales_channel_i)
  • Constraints:
    • Total Spend <= Budget
    • Min_Spend_i <= Spend_i <= Max_Spend_i (Business constraints, e.g., “We must spend at least $50k on Brand Search”).
  • Algorithm: We utilized scipy.optimize with the L-BFGS-B (Limited-memory Broyden–Fletcher–Goldfarb–Shanno Boxed) algorithm, which is highly efficient for bound-constrained non-linear optimization problems.

3. Scenario Planner

We built a web-based “Simulator” tool.

  • Input: “What if I have $5M? What if I have $10M?”
  • Output: The tool visualizes the optimal split. It also shows the “Opportunity Cost” (how much revenue you lose by deviating from the optimal).

Implementation Details

  • Tech Stack: Python, Pandas, SciPy, Streamlit (for the UI).
  • Scalability: Validated the optimizer on up to 50 distinct channels/geographies.
  • Integration: Outputs were exported to Excel/Tableau for the Media Buying teams to execute.

Challenges & Solutions

  • Challenge: Local Optima. The curve might be complex.

  • Solution: “Basinhopping” (Global Optimization). We ran the optimizer from multiple random starting points to ensure we found the global maximum, not just a local peak.

  • Challenge: “The model says spend $0 on TV.” (Political Challenge).
  • Solution: Hard Constraints. We added capability for stakeholders to lock specific channels (“Force TV spend = $1M”), and the optimizer would find the best possible solution given that constraint, while highlighting the efficiency loss.

Results and Impact

  • ROI Improvement: The optimized allocation was projected to deliver 15-20% more revenue for the same total budget compared to the historical allocation.
  • Adoption: The tool became the standard quarterly planning instrument for the CMO office.
  • Efficiency: Automated a process that previously took 2 weeks of Excel work into a 5-minute run.

Future Work

  • Multi-Objective Optimization: Optimizing for both Short-term Revenue AND Long-term Brand Equity (Awareness) simultaneously (Pareto Frontier).
  • Granularity: Moving from channel-level (TV vs Digital) to campaign-level optimization.