Diego S. · Bayesian Media Mix Model

Portfolio · Case Study

Diego S.

Marketing Analytics Bayesian Modeling Media Mix Modeling Data Science

Education

Florida International University

BS · Computer Science

Florida International University

MBA · Marketing & E-Commerce

Pennsylvania State University

Master of Applied Statistics

Available now

Expert-Vetted

Top Rated

100%Job Success

Case Study

Marketing
Effectiveness Platform

Bayesian MMM + BI Dashboards

Insurance

Marketing Analytics

$171M

Media spend analyzed

5

Media channels optimized

9

Data sources unified

01 · 07

Project overview

Key project facts

Client

Fortune 100 Insurance Provider

Industry

Insurance · Financial Services

Engagement type

Marketing Analytics

Approach

Bayesian MMM (PyMC-Marketing)

Scope

B2C segment · 5 channels

Technical stack

Python PyMC-Marketing statsmodels R Azure SQL Server Tableau

Engagement summary

A marketing effectiveness engagement for a Fortune 100 insurance provider, spanning two complementary deliverables. Nine marketing and business platforms — media cost, web analytics, email, CRM, call center, and policy management — were unified into a single weekly time series via an automated Python/R pipeline. On top of that dataset came both deliverables: an interactive Tableau dashboard suite — live-querying the database so marketing leadership could explore spend, leads, and conversions across segments on demand — and a Bayesian Media Mix Model to quantify what was actually driving B2C policy volume. The model, built on PyMC-Marketing, applies adstock (carryover) and saturation transforms to each channel, then fits channel contributions with a PyMC NUTS MCMC sampler — returning full posterior distributions, not point estimates: per-channel ROI with credibility intervals, contribution shares, response curves, and a budget-reallocation recommendation. The Bayesian approach captures the delayed and diminishing media effects — carryover and saturation — that the original linear regression could not.

Bayesian

Posterior Inference

Full posterior distributions per channel — ROI with credibility intervals, not point estimates

$171M

Media Spend Analyzed

Full platform across all segments; MMM scoped to B2C

5

Channels Modeled

TV · Paid Search · Display · Print · Video — each with adstock + saturation

02 · 07

Technical Design · Media Mix Model

Data & Feature Pipeline

9 Sources · Weekly Aggregation · B2C Scope

Sources

Mktg Cloud Intelligence

media cost · impressions

Google Analytics

web traffic

SLMS

policy management

Sutherland

call center

Mktg Cloud Email Studio

email · +CRM, TV, print, display = 9 total

ETL

Python + R

automated pipeline

W-SAT weekly

Saturday week-ending alignment

Channel pivot

spend → per-channel columns

Cleansing

type casts · name standardization

Target Construction

leads

raw inquiries

→

policies

converted

→

premiums

revenue

monthly ×1.5 annual ×12 LTV ×4 (assumption) groupby week

Extra Features

Holiday dummies

holidays.USA

↓

Seasonal decompose

statsmodels

↓

Conversion rate

leads ÷ policies

Matrix

Media matrix

5 channels weekly

Extra features

holidays + seasonality

Target

policies + premiums

~2 yrs

weekly granularity

Channels

TV

largest budget

Paid Search

Display

Print

Video

Technical Approach

Nine marketing and business platforms were unified into a single weekly time series. Media spend, leads, policies, and premiums were all aligned to a Saturday week-ending boundary so media activity and conversion outcomes share the same calendar. The target (policy volume) and a set of extra features — US holidays and seasonal components — were assembled into a model-ready matrix scoped to the B2C segment.

Key Decisions

W-SAT weekly alignment — media spend and lead/policy outcomes are bucketed to the same Saturday week-ending so the model compares like-for-like time periods.

B2C segment scope — the model is built for one business segment, not pooled across B2C, B2B, Agent, and Broker. Media-to-conversion dynamics differ too much to pool.

Holiday + seasonal features — calendar effects are modeled explicitly so media coefficients aren't contaminated by seasonality.

03 · 07

Technical Design · Media Mix Model

Bayesian Media Mix Model

PyMC-Marketing · NUTS MCMC · Adstock + Saturation

1

Inputs

Media + Features + Target

5 media channels (weekly spend) · extra features (holidays, seasonality) · target: B2C policy volume · all scaled

weekly matrix

2

Media Transform

Adstock + Hill saturation

Carryover (adstock) models delayed media effect over time · Hill saturation models diminishing returns per channel · captures the lag direct attribution misses

carryover + saturation

3

Inference

PyMC-Marketing · PyMC

NUTS MCMC sampler · tune + draws · channel priors · posterior distributions over channel effects, not point estimates

NUTS MCMC

Sampler config

NUTS warmup + samples channel priors yearly seasonality

4

Diagnostics

r_hat Convergence

Gelman-Rubin r_hat checked across chains · values near 1.0 confirm the sampler converged before any inference is trusted

r_hat ≈ 1.0

5

Model Outputs

Channel Posteriors

ROI per channel with 0.05–0.95 credibility intervals · media contribution % · response/saturation curves · baseline vs. media decomposition

credibility intervals

6

Recommendation

budget optimizer

Constrained optimization reallocates the fixed budget across channels to maximize predicted policy volume · pre/post allocation comparison delivered to marketing leadership

budget reallocation

Why Bayesian MMM

A Bayesian model returns full posterior distributions over each channel's effect — not single point estimates. That means every ROI figure comes with a credibility interval, so leadership sees the uncertainty, not just a number. Adstock and saturation transforms let the model capture delayed and diminishing media effects that linear attribution cannot.

Key Decisions

PyMC-Marketing over linear regression — the original engagement used multiple linear regression. The Bayesian approach adds adstock and saturation, modeling carryover and diminishing returns that a linear model treats as instantaneous and unbounded.

Adstock captures carryover — TV's effect persists for days after a flight. Adstock models this delay, which is why TV shows the highest contribution despite the lowest per-dollar direct ROI.

Credibility intervals over point estimates — budget decisions are made with the uncertainty visible, not hidden behind a single ROI number.

04 · 07

Model findings

TV drives the most volume — at the lowest per-dollar ROI

TV was not underperforming. Its return was delayed, and direct attribution couldn't see it.

The Bayesian model estimates each channel's media contribution and its per-dollar ROI separately. TV shows the highest media contribution of any channel — yet the lowest per-dollar ROI. That tension is the whole story: TV's effect is spread out over time through adstock carryover, so a direct, same-period attribution model credits it with almost nothing. The MMM's adstock transform recovers that delayed contribution, quantifying the awareness TV generates that converts later through other channels.

Before

Direct attribution credited paid search with conversions and TV with almost none. TV — the largest budget line — appeared to underperform, with no way to measure its delayed effect.

After

Adstock carryover quantifies TV's delayed contribution. Leadership sees channel contribution and per-dollar ROI as two separate truths — and optimizes the awareness-to-capture spend ratio, not just last-touch ROI.

Highest

TV media contribution

Largest share of modeled policy volume of any channel — recovered through adstock carryover

Lowest

TV per-dollar ROI

Direct ROI understates TV's true value because its effect is delayed, not immediate

Adstock

Carryover modeled

Adstock + saturation transforms capture the delayed media effect a linear, same-period model treats as zero

05 · 07

Illustrative model outputs

What the model produces

Model Fit: Predicted vs. Actual

Bayesian posterior prediction tracks weekly policy volume, with credibility band

Channel ROI with Credibility Intervals

Per-dollar ROI per channel with Bayesian credibility intervals — paid search highest, TV lowest

Media Contribution by Channel

TV drives the largest share of modeled volume despite the lowest per-dollar ROI — the carryover effect

Response & Saturation Curves

Diminishing-returns curve per channel — the marginal policy volume earned at each spend level

06 · 07

Deliverables

BI dashboard views

All Segments Overview

Executive view — media cost, impressions, converted leads, and annualized premium by segment and channel

Media Mix vs. Cost Per Lead

Monthly spend by channel overlaid with CPL trend — segmentable by B2C, B2B, Agent, Broker

Customer Lifecycle Funnel

Bubble KPIs mapping media cost through impressions, clicks, and form completions to converted leads

Trend & Correlation Analysis

Any two KPIs compared as a dual-axis time series and correlation scatterplot, configurable by segment

07 · 07