Marketing Mix Modeling Platform

Overview

Engineered a marketing mix modeling platform using BayesianRidge regression on STL-decomposed residuals. Built a parallelized Hill Curve Transformer (L-BFGS-B multi-start optimization) for saturation modeling at 1000+ feature scale, adstock decay features, PCA-based group-level attribution, and response curve generation (isotonic + spline/hill variants). Implemented scenario planning via response-curve delta algorithm with business-rule guardrails to prevent cascade drift. Fixed critical production bugs: MinMax extrapolation bug, scenario planner 6.8% cascade drift, STL phantom baseline edge cases.

The problem

Marketing teams had spend across multiple channels but no systematic way to measure each channel's contribution to pipeline generation or optimize budget allocation. Rule-based attribution couldn't capture saturation effects, adstock carryover, or the non-linear relationship between spend and return.

How the system works

1. 01

   Spend and outcome data

   by channel

2. 02

   Transforms

   adstock decay and Hill saturation curves

3. 03

   Seasonal decomposition

   STL

4. 04

   Response curves

   per channel

5. 05

   Scenario planning

   budget reallocation

What broke

• MinMax extrapolation
• Scenario cascade drift
• STL edge cases

Results

Quantified channel-level marketing contribution and enabled spend optimization recommendations for enterprise B2B SaaS customers. Analysis guided reallocation decisions, e.g., identifying that reducing spend on one underperforming channel while maintaining pipeline output was viable.