Rituraj's Garden

Graph View

Curriculum - Marketing Mix Modeling

16 min read

Executive Overview

Domain: Marketing Science & Econometrics for Marketing Measurement Role Focus: Data Lead responsible for production MMM, experimentation, optimization, and governance Timeline: 16-40 weeks (depending on depth) Prerequisites: Strong data engineering background, python/R proficiency, basic statistics

Your Core Deliverables as Data Lead

As the Data Lead for MMM in your CPG company, you must own:

  1. Production MMM System - Automated, quarterly-refreshed models using Python or R
  2. Experiment Design Playbooks - GeoLift templates and protocols
  3. Budget Optimizer - Constraint-based allocation engine
  4. KPI Framework - Unified measurement across MMM, experiments, clean rooms
  5. Governance Documentation - Model cards, data lineage, assumption logs
  6. Education Deck - Executive-level explainer for non-technical stakeholders
  7. Measurement Stack Policy - How MMM, experiments, and clean rooms integrate

Curriculum Structure

Quick Track (12-16 weeks): Production-focused

Covers core theory, two implementations (Robyn + PyMC-Marketing), validation, and deployment

Deep Track (32-40 weeks): Expert mastery

Adds advanced statistics, CPG specialization, cutting-edge research, and organizational leadership

QUICK TRACK: Production-Ready MMM (12-16 Weeks)

Week 0: Orientation & Scope (2 days)

Objectives

  • Understand MMM’s role in modern marketing measurement
  • Grasp the “four stacks”: theory, modeling, experimentation, production
  • Map your learning journey

Essential Reading

  • “An Analyst’s Guide to MMM” (Meta Robyn) - facebookexperimental.github.io
  • “Marketing Mix Modeling: A Complete Guide for Strategic Marketers” (Measured, 2025)
  • “Why marketing mix modeling is crucial in 2025 and beyond” - Search Engine Land

Key Concepts

  • MMM vs Attribution vs Incrementality testing
  • Privacy-first measurement landscape (post-cookie era)
  • MMM’s 2025 renaissance: why now?
  • CPG-specific considerations

Action Items

  • Read Analyst’s Guide end-to-end
  • Identify your company’s current measurement gaps
  • List available data sources
  • Define success metrics for this learning journey

Week 1: Market Response Foundations

Core Theory: Demand Drivers & Market Response

Key Concepts:

  • Base sales vs. incremental sales
  • Endogenous vs. exogenous variables
  • Competitive effects and market dynamics
  • Seasonality patterns in CPG
  • Promotional mechanics: BOGO, TPR, price discounts
  • Distribution and availability effects

Essential Reading

Core Textbook (chapters 1-3, 6):

  • “Market Response Models” (2nd ed.) by Hanssens, Parsons & Schultz - SpringerLink
    • The foundational text for market response modeling
    • Covers demand theory, elasticity, and dynamics

Survey Papers:

  • “Market Response Models and Marketing Practice” - Hanssens et al. - ResearchGate

    • Bridges theory and practice
    • Shows what actually gets used in industry

  • “Modeling Marketing Dynamics by Time Series Econometrics” - Dekimpe & Hanssens - SpringerLink

    • Time-series perspective on marketing effects
    • Dynamic modeling frameworks

  • “Time-Series Models in Marketing” - Naik et al. - UC Davis

    • Compact primer on time-series methods
    • Practical examples

CPG-Specific Context

  • Product life cycles and category dynamics
  • Retailer power and trade spending
  • Regional market differences
  • Competitive response patterns
  • Forward buying and pantry loading

Practical Exercises

  • Decompose historical sales into base + incremental
  • Calculate price and promotional elasticities
  • Identify structural breaks in your data
  • Map competitor activity impact

Action Items

  • Read Hanssens chapters 1-3, 6
  • Work through elasticity calculations
  • Document your CPG category’s unique dynamics
  • Create glossary of CPG promotional terms

Week 2: Adstock & Saturation Mechanics

Core Theory: Non-Linear Response

Adstock (Carryover Effects):

  • Memory effect of advertising
  • Decay over time after exposure
  • Channel-specific persistence

Saturation (Diminishing Returns):

  • S-curve response to marketing spend
  • Threshold and ceiling effects
  • Optimal spending levels

Essential Reading

Adstock:

  • “Adstock revisited” - Franses - Erasmus University

    • Modern treatment of Koyck lag structure
    • Default to unrestricted Koyck unless data prove otherwise

  • “Understanding the Role of Adstock in Advertising Decisions” - Gijsenberg et al.

    • Strategic implications of adstock
    • Business decision-making

  • “What are adstock and carryover effects and how can they be modelled?” - Sellforte

    • Practical implementation guide
    • Different transformation approaches

Saturation:

  • “Media saturation and lagging” - Google Meridian docs - Google for Developers

    • Hill function formalism
    • Prefer Hill over ad-hoc splines for stability

  • “Exploring Different Approaches to Generate Response Curves in MMM” - Medium

    • Comparison of saturation functions
    • Visual intuition

Key Mathematical Concepts

Geometric Adstock:

At = Tt + λ * At-1
where λ ∈ (0, 0.95) is decay parameter

Delayed Geometric:

Allows peak effect after lag period

Weibull Adstock:

Flexible shape parameter
Used in advanced implementations

Hill Saturation:

y = α * x^s / (K^s + x^s)
where:
- α = ceiling (maximum response)
- s = shape (steepness)
- K = half-saturation point

Typical Adstock Decay by Channel

(from Meta Robyn guidelines)

  • TV: 0.3–0.8
  • OOH/Print/Radio: 0.1–0.4
  • Digital: 0.0–0.3
  • Social Media: 0.0–0.3
  • Search: 0.0–0.1 (very short decay)

Practical Exercises

  • Implement geometric adstock transformation
  • Fit Hill saturation curves to simulated data
  • Compare different decay rates visually
  • Calculate half-life of advertising effects

Action Items

  • Read Franses “Adstock revisited”
  • Read Google Meridian saturation docs
  • Code adstock transformations in Python
  • Visualize response curves for your channels
  • Estimate appropriate decay ranges for your media

Weeks 3-4: Core MMM Modeling

Frequentist vs. Bayesian Approaches

Frequentist (Traditional):

  • Point estimates only
  • No uncertainty quantification
  • Faster for simple models
  • Less flexible

Bayesian (Modern Default):

  • Full posterior distributions
  • Natural uncertainty quantification
  • Incorporates prior knowledge
  • Better for optimization under uncertainty
  • Use this for production systems

Three Production-Ready Frameworks

1. Meta Robyn (R-based)

Characteristics:

  • Ridge regression + multi-objective optimization
  • Hyperparameter search via Nevergrad
  • Prophet for trend/seasonality decomposition
  • Production-hardened, strong diagnostics
  • Automated model selection
  • Built-in budget allocator

Key Features:

  • Multi-touch attribution integration
  • GeoLift calibration support
  • Response curve visualization
  • Pareto frontier optimization

When to Use:

  • Quick production deployment
  • Team familiar with R
  • Need automated workflows
  • Want proven, battle-tested solution

Essential Resources:

2. Google LightweightMMM (python)

Characteristics:

  • Bayesian via PyMC/NumPyro
  • MCMC sampling (NUTS)
  • Explicit prior specifications
  • Clean, minimal API

Key Features:

  • Hierarchical geo-level modeling
  • Posterior predictive checks
  • Adstock + saturation built-in
  • Good for learning Bayesian concepts

When to Use:

  • Python environment
  • Want full Bayesian treatment
  • Building custom models
  • Educational/research purposes

Status: Note that Google Meridian is the successor (released 2024), but LightweightMMM still valuable for learning

Essential Resources:

  • LightweightMMM GitHub - github.com/google/lightweight_mmm
  • Code walkthrough and notebooks
  • “Understanding Bayesian Marketing Mix Modeling” - Towards Data Science

3. PyMC-Marketing (python)

Characteristics:

  • Most flexible Bayesian framework
  • Built on PyMC ecosystem
  • Time-varying coefficients
  • Advanced diagnostics

Key Features:

  • Complete MMM module with budget optimizer
  • Causal inference wrappers
  • Extensive posterior analysis tools
  • Active development, strong community
  • Integration with ArviZ for diagnostics

When to Use:

  • Need maximum flexibility
  • Custom model requirements
  • Advanced Bayesian modeling
  • End-to-end Bayesian pipeline in python

Essential Resources:

  • PyMC-Marketing Documentation - pymc-marketing.io
  • MMM example notebook - Link
  • Budget optimization API documentation
  • PyMC Discourse community

Learning Path for Weeks 3-4

Week 3: Robyn Deep Dive

  • Install and setup Robyn
  • Work through official tutorials
  • Run on demo dataset
  • Understand hyperparameter optimization
  • Generate response curves and budget allocation

Week 4: Python Bayesian MMM

  • Setup LightweightMMM or PyMC-Marketing
  • Understand prior specification
  • Run MCMC sampling
  • Perform posterior predictive checks
  • Compare with Robyn results

Practical Exercises

Exercise 1: Robyn Implementation

# Setup, data prep, model run
# Hyperparameter optimization
# Model selection via Pareto frontier
# Response curve analysis
# Budget optimization

Exercise 2: PyMC-Marketing Implementation

# Prior specification
# Model definition with adstock + saturation
# MCMC sampling with diagnostics
# Posterior analysis
# Budget optimization under constraints

Key Concepts to Master

  • Bayesian priors and posteriors
  • MCMC convergence diagnostics (R-hat, ESS)
  • Hyperparameter vs. model parameters
  • Model selection criteria (Pareto optimality, DIC, WAIC)
  • Credible intervals vs. confidence intervals

Action Items

  • Install Robyn and dependencies
  • Complete Robyn quickstart tutorial
  • Install PyMC-Marketing
  • Run both frameworks on same demo data
  • Document differences in outputs
  • Create comparison matrix of frameworks

Weeks 5-6: Causal Identification & Incrementality

The Observational Problem

Key Challenge: MMM uses observational data, not randomized experiments

  • Correlation ≠ causation
  • Confounding variables
  • Endogeneity issues
  • Selection bias

Solution: Pair MMM with experiments to:

  1. Pin down priors (Bayesian calibration)
  2. Validate elasticity estimates
  3. Ground-truth lift measurements
  4. Build stakeholder confidence

Geo-Based Experiments

Meta GeoLift: Synthetic Control at Geo Level

What It Is:

  • Synthetic control method for geo experiments
  • Matched market testing
  • Power calculations pre-launch
  • Post-campaign lift measurement

Key Concepts:

  • Treatment vs. control geos
  • Synthetic control weighting
  • Minimum detectable lift (MDL)
  • Pre-period validation

Use Cases in CPG:

  • TV market testing
  • Retail media incrementality
  • Regional promotional lifts
  • Distribution expansion tests

Essential Resources:

Bayesian Structural Time Series (BSTS)

What It Is:

  • CausalImpact framework
  • Time-series counterfactual estimation
  • Useful for single-treatment lifts

Use Cases:

  • Campaign launch impact
  • Channel pause tests
  • Budget shift experiments

Essential Resources:

  • PyMC-Marketing BSTS implementation - Docs
  • CausalImpact R package (Google)
  • “Inferring causal impact using Bayesian structural time-series models”

Privacy and Signal Loss Context

Why MMM Matters More in 2025:

  • ATT (App Tracking Transparency) impact
  • Third-party cookie deprecation
  • GDPR/CCPA regulations
  • Walled garden fragmentation

Implications:

  • Attribution models breaking down
  • User-level tracking disappearing
  • MMM as privacy-safe alternative
  • Aggregate-level measurement rising

Essential Reading:

  • “Marketing & Measuring In Light of iOS Privacy” - Prescient AI
  • “The comeback of the marketing mix model” (2025 trends)
  • eMarketer reports on measurement shifts

Integrating Experiments with MMM

Workflow:

  1. Design geo experiment using GeoLift power calculations
  2. Run experiment in treatment vs. control markets
  3. Measure lift with synthetic control
  4. Calibrate MMM priors using experiment results
  5. Validate MMM elasticities against experiments
  6. Iterate quarterly with new experiments

Calibration Methods:

  • Use experiment lifts as priors in Bayesian MMM
  • Constrain effect sizes to match experiments
  • Weight recent experiments more heavily
  • Document calibration in model governance

Practical Exercises

Exercise 1: GeoLift Power Analysis

  • Define treatment/control geos
  • Calculate minimum detectable lift
  • Run synthetic control matching
  • Validate pre-period fit

Exercise 2: MMM Calibration

  • Take experiment lift result
  • Specify Bayesian prior around that lift
  • Re-run MMM with calibrated priors
  • Compare uncalibrated vs. calibrated

Exercise 3: Validation Framework

  • Design quarterly experiment calendar
  • Map experiments to MMM channels
  • Create experiment registry
  • Build validation dashboard

Action Items

  • Complete GeoLift tutorial
  • Design one geo experiment for your business
  • Calculate power and MDL
  • Read privacy measurement papers
  • Create experiment-MMM integration plan
  • Draft experiment playbook template

Week 7: Retail Media Networks & Clean Rooms for CPG

The Retail Media Revolution

What Are Retail Media Networks (RMNs)?:

  • Advertising on retailer platforms (Amazon, Walmart, Target)
  • First-party retailer data for targeting
  • Closed-loop measurement (purchase data)
  • Fastest-growing ad channel in CPG

Key RMNs for CPG:

  • Amazon Advertising (dominant)
  • Walmart Connect
  • Target Roundel
  • Kroger Precision Marketing
  • Instacart Ads

Why RMNs Matter for MMM

Challenges:

  • Walled gardens with limited data export
  • Inconsistent measurement standards
  • Attribution complexity
  • Data integration difficulties

2025 Trend: RMNs adding MMM-ready data feeds

  • Standardized exports
  • Geo-level aggregation
  • Privacy-compliant feeds
  • Direct API access

Essential Reading:

  • “2025 trend: Retail media networks wade into unfamiliar waters” - eMarketer
  • “2025 Retail Media Market Guide” - RMIQ

Amazon Marketing Cloud (AMC)

What It Is:

  • Clean room solution on AWS Clean Rooms
  • sql-based analysis environment
  • Aggregated, privacy-safe queries
  • Cross-channel attribution

Key Capabilities:

  • Path-to-purchase analysis
  • Audience overlap studies
  • Frequency and reach
  • Incrementality measurement
  • Integration with non-Amazon data

Setup Requirements:

  • AMC instance provisioning
  • sql query expertise
  • Data schema understanding
  • Privacy compliance knowledge

Essential Resources:

  • Amazon Marketing Cloud Documentation - Amazon Ads
  • AMC certification program
  • SQL query templates
  • Clean room concepts guide

Data Integration Strategy

What You Must Ingest:

  1. Amazon Ads:

    • Sponsored Products/Brands/Display spend
    • Impressions, clicks, conversions
    • ACOS, ROAS metrics
    • Geo-level if available

  2. Other RMNs:

    • Standardized metrics across platforms
    • Weekly or daily granularity
    • Campaign-level details

  3. Clean Room Outputs:

    • AMC query results
    • Aggregated audience insights
    • Attribution weights

Harmonization Challenges:

  • Different metrics definitions
  • Varying time zones
  • Attribution window differences
  • Geo granularity mismatches

Building RMN-Ready MMM

Data Pipeline Requirements:

  • Automated daily/weekly pulls
  • Transformation to MMM grain (weekly)
  • data quality checks for data validation
  • Version-controlled processing

Modeling Considerations:

  • Separate RMN channels from other digital
  • Account for closed-loop nature
  • Model retailer-specific effects
  • Include distribution as control

Governance:

  • Data sharing agreements
  • Privacy compliance checks
  • Access controls
  • Audit trails

Practical Exercises

Exercise 1: AMC Query Building

  • Write sql for path analysis
  • Calculate reach/frequency
  • Export for MMM integration

Exercise 2: RMN Data Pipeline

  • Design etl for multiple RMNs
  • Harmonize to common schema
  • Create QA dashboard

Exercise 3: RMN MMM Module

  • Add RMN channels to existing MMM
  • Model retailer-specific effects
  • Optimize RMN budget allocation

Action Items

  • Map current RMN investments
  • Request data access from RMN partners
  • Complete AMC certification
  • Build RMN data integration pipeline
  • Design RMN measurement strategy
  • Document RMN-specific model considerations

Week 8: Optimization & Decisioning

Budget Allocation Under Uncertainty

Core Principle: Use Bayesian posterior draws to optimize expected profit subject to constraints

Why Bayesian?:

  • Incorporates uncertainty naturally
  • Multiple scenarios from posterior
  • Risk-adjusted decisions
  • Credible intervals for recommendations

Optimization Frameworks

Response Curve Optimization

Concept: Find budget allocation that maximizes response along fitted curves

Methods:

  1. Gradient-based (when curves are smooth)
  2. Genetic algorithms (Robyn approach)
  3. Bayesian optimization (when expensive to evaluate)
  4. Linear programming (convex cases)

Key Metrics

Average ROI vs. Marginal ROI:

  • Average ROI: Total return / total spend
  • Marginal ROI: Additional return from next dollar
  • Optimize on marginal, report average

Expected Profit:

E[Profit] = E[Revenue] - Total Spend - Fixed Costs

Risk Metrics:

  • Value at Risk (VaR)
  • Expected shortfall
  • Probability of meeting targets

Constraint Types

Business Constraints:

  1. Total Budget: $100K per quarter
  2. Channel Minimums: At least $5K per channel
  3. Channel Maximums: No more than 40% in any channel
  4. Change Limits: ±30% vs. historical spend
  5. Flight Windows: Minimum run periods
  6. Geo Requirements: Regional balance

Strategic Constraints:

  • Brand vs. performance balance
  • Long-term vs. short-term split
  • Test budget allocation (10-15%)
  • Competitive response limits

Implementation: Robyn Budget Allocator

How It Works:

  • Takes calibrated MMM results
  • Simulates budget scenarios
  • Optimizes via Nevergrad
  • Returns Pareto-optimal allocations

Two Scenarios:

  1. Max Response: Given budget, maximize KPI
  2. Target Efficiency: Reach target with minimum spend

Key Outputs:

  • Optimal spend by channel
  • Expected lift vs. baseline
  • Response curves with current/optimal points
  • Decomposition of effects

Usage:

robyn_allocator(
  robyn_object = robyn_model,
  scenario = "max_response",
  total_budget = 100000,
  channel_constr_low = 0.7,  # -30%
  channel_constr_up = 1.3     # +30%
)

Essential Resource:

Implementation: PyMC-Marketing Budget Optimizer

How It Works:

  • Uses posterior draws from MMM
  • Constraint-based optimization
  • Multiple objectives supported
  • Uncertainty propagation

Key Features:

  • Custom objective functions
  • Flexible constraints
  • Scenario comparison
  • Risk analysis

Usage:

from pymc_marketing.mmm import BudgetOptimizer
 
optimizer = BudgetOptimizer(mmm_model)
optimal_allocation = optimizer.optimize(
    total_budget=100000,
    constraints={
        "tv": (0.7, 1.3),
        "digital": (0.7, 1.3)
    }
)

Essential Resources:

Scenario Planning Workflow

Step 1: Define Scenarios

  • Base Case: Current spend allocation
  • Optimized: Algorithm recommendation
  • Conservative: ±15% changes only
  • Aggressive: ±50% changes allowed
  • What-If: Specific strategic shifts

Step 2: Run Simulations

  • Pull N samples from posterior (e.g., 1000)
  • Optimize each sample
  • Aggregate results

Step 3: Risk Analysis

  • Calculate percentiles (10th, 50th, 90th)
  • Identify downside risks
  • Assess probability of meeting targets

Step 4: Create Decision Packages

  • Multiple options for CFO
  • Risk-return tradeoffs
  • Sensitivity analysis
  • Downside protection

Executive-Ready Outputs

CFO Memo Template:

  1. Executive Summary: Recommendation in 3 bullets
  2. Current Performance: Historical ROI by channel
  3. Optimization Results: Table of scenarios
  4. Expected Impact: Revenue lift with credible intervals
  5. Risk Assessment: Downside scenarios
  6. Implementation Plan: Timeline and dependencies
  7. Validation: How we’ll measure success

Visualization Best Practices:

  • Response curves with current/optimal markers
  • Waterfall chart of reallocation
  • Scenario comparison table
  • Risk distributions (violin plots)
  • Before/after allocation pie charts

Practical Exercises

Exercise 1: Robyn Optimization

  • Run allocator on demo model
  • Compare max response vs. target scenarios
  • Vary constraints and observe changes
  • Document optimal allocation

Exercise 2: PyMC-Marketing Optimization

  • Implement custom objective function
  • Add business constraints
  • Run scenario analysis
  • Visualize uncertainty

Exercise 3: CFO Decision Package

  • Create 3 scenarios with different risk profiles
  • Calculate expected profit and downside risk
  • Build executive presentation
  • Practice stakeholder communication

Action Items

  • Implement both Robyn and PyMC-Marketing optimizers
  • Define business constraints for your company
  • Build scenario planning framework
  • Create CFO memo template
  • Design risk analysis dashboard
  • Document optimization assumptions and limitations

Week 9: CPG Data Engineering for MMM

Data Requirements: The Complete Stack

Core Principle: Align all data to weekly grain with consistent lagging, QA checks, and gold-standard tables

Required Data Tables

1. Sales Data (Dependent Variable)

Weekly Sales by Brand/SKU:

  • Revenue or units sold
  • Multiple products/SKUs if portfolio
  • Regional splits if modeling hierarchically
  • Retailer-level if available (and strategic)

Sources:

  • Internal sales systems
  • Nielsen/IRI scanner data (stored in data warehouse)
  • Retailer POS data
  • Circana (formerly IRI) data

Schema (data modeling):

week_start_date | brand | sku | region | retailer | units | revenue | market_share

2. Media Spend & Exposure

Digital Channels:

  • Spend, impressions, clicks by week
  • Platform: Facebook, Google, TikTok, etc.
  • Campaign/ad set level detail
  • Geographic splits if available

Traditional Channels:

  • TV: Spend + GRPs (Gross Rating Points)
  • Radio: Spend + GRPs
  • OOH (Out of Home): Spend + impressions
  • Print: Spend + circulation

Retail Media:

  • Amazon Ads (Sponsored Products/Brands/Display)
  • Walmart Connect
  • Other RMNs
  • Separate from general digital

Schema:

week_start_date | channel | campaign | spend | impressions | clicks | region

3. Price & Promotions

Base Price:

  • Regular shelf price by SKU/week
  • Account for inflation if multi-year

Promotional Price:

  • Discounted price during promotions
  • Promotion type: TPR, BOGO, % off, etc.
  • Promotion depth: % discount
  • Promotional timing and duration

Price Indices:

  • Own price relative to base
  • Competitive price indices
  • Price gaps vs. competitors

Schema (data modeling):

week_start_date | sku | region | base_price | promo_price | promo_type | promo_depth | promo_flag

4. Distribution & Availability

Weighted Distribution:

  • % of stores carrying product
  • Weighted by store volume (ACV)
  • Critical for new product launches

Numeric Distribution:

  • Simple count of stores
  • Less preferred than weighted

Out-of-Stock Rates:

  • Availability at shelf
  • Impacts sales directly

Schema (data modeling):

week_start_date | sku | region | weighted_dist | numeric_dist | oos_rate

5. Trade Spend

Co-op Advertising:

  • Retailer-funded marketing
  • Must separate from brand spend

Slotting Fees:

  • One-time payments for shelf space
  • Amortize over relevant period

Other Trade:

  • Display fees
  • Promotional allowances

Schema (data modeling):

week_start_date | retailer | trade_type | amount

6. Competitor Data

Competitive Spend:

  • Estimated via Kantar or similar
  • Share of voice (SOV)

Competitive Pricing:

  • Shelf prices of key competitors
  • Price indices

Competitive Promotions:

  • Promotion frequency and depth
  • Timing of major campaigns

Schema:

week_start_date | competitor | channel | est_spend | price_index | promo_flag

7. Seasonality & Calendar

Holidays:

  • National holidays
  • Religious holidays (regional)
  • School breaks
  • Major shopping events (Black Friday, Prime Day)

Events:

  • Product launches
  • PR events
  • Sponsorships
  • Corporate events

Day of Week Effects:

  • If modeling daily, not weekly

Schema:

week_start_date | holiday_name | event_name | holiday_flag | event_flag

8. Macro & External Factors

Economic Indicators:

  • Consumer confidence index
  • Unemployment rate
  • Gas prices (for some CPG categories)
  • Inflation rates

Weather:

  • Temperature by region
  • Precipitation
  • Extreme events
  • Relevant for seasonal products

COVID/Pandemic:

  • Lockdown indicators
  • Mobility indices
  • Category-specific shocks

Schema:

week_start_date | region | temp_avg | precip | unemployment_rate | consumer_confidence

Data Engineering Best Practices

Gold Table Architecture

Principle: One source of truth for MMM inputs

Structure:

mmm_gold_weekly
├── week_start_date (Monday)
├── sales_revenue
├── sales_units
├── tv_spend
├── tv_grps
├── digital_spend
├── digital_impressions
├── ... (all media channels)
├── base_price
├── promo_depth
├── promo_flag_bogo
├── promo_flag_pctoff
├── weighted_distribution
├── comp_sov
├── holiday_flag
├── temperature
└── ... (all control variables)

Requirements:

  • Weekly grain (Monday start recommended)
  • No nulls (forward-fill or interpolate)
  • Consistent units (don’t mix thousands and millions)
  • Documented transformations with data governance
  • Version controlled with data quality checks

Lagged Variables & Adstock

Pre-compute Adstock (python):

def geometric_adstock(x, decay):
    """Apply geometric adstock transformation"""
    adstocked = np.zeros_like(x)
    adstocked[0] = x[0]
    for i in range(1, len(x)):
        adstocked[i] = x[i] + decay * adstocked[i-1]
    return adstocked
 
# Create adstocked versions
df['tv_adstock_30'] = geometric_adstock(df['tv_spend'], 0.3)
df['tv_adstock_50'] = geometric_adstock(df['tv_spend'], 0.5)
df['tv_adstock_70'] = geometric_adstock(df['tv_spend'], 0.7)

Recent Posts

Recent Notes