What it actually does vs. what people think it does
The Brutal Truth in 30 Seconds
Granger Causality: A test that checks if X happens before Y and helps predict it.
That's it. Not causation. Not magic. Just temporal correlation with extra steps.
✅ What Granger ACTUALLY Does
Tests if X's past helps predict Y's future
Finds temporal ordering (what comes first)
Gives you a p-value (yes/no answer)
Works for linear relationships
❌ What Granger DOESN'T Do
Prove actual causation
Tell you effect size (HOW MUCH)
Handle non-linear relationships well
Work with small samples (<100 obs)
The 95/5 Reality Check
95% of Marketing Questions:
"HOW MUCH does TV drive sales?" (Need effect size, not yes/no)
5% Where Granger Might Help:
"Does weather affect sales AT ALL?"
The Rooster and Sunrise Problem
🐓 Classic Granger Fallacy:
Rooster crows → Sunrise (Granger says: ✓ Significant!)
Does rooster CAUSE sunrise? Obviously not.
🛍️ Marketing Equivalent:
Google searches → Purchases (Granger says: ✓ Significant!)
Do searches CAUSE purchases? Maybe not - people search AFTER deciding to buy.
Better Alternatives for Every Use Case
What You're Trying to Do
Granger Approach
Better Alternative
Why It's Better
Find lag structure
Test significance at each lag
Cross-correlation function
Shows magnitude, not just yes/no
Select variables
Test each variable separately
LASSO/Ridge regression
Handles all variables at once
Prove causation
Show temporal precedence
Run an experiment
Actually proves causation
Measure ROI
Can't do this at all
Incrementality test
Gives you actual $$$ impact
Competitive dynamics
Test who leads whom
Actually useful here!
One of few good use cases
When People Actually Use Granger (The Honest Version)
# How Granger selection really happens in practice:
1. Include everything client thinks matters ✓
2. Add competitors if data available ✓
3. Add seasonality and holidays ✓
4. Let regularization kill weak variables ✓
5. Check business logic on survivors ✓
6. "Hmm, should weather be in here?"
7. Run Granger test
8. Include anyway because CEO insists 🤷
The One-Sentence Summary
Granger causality is the appendix of statistical methods - evolution left it there,
it occasionally gets inflamed and causes problems, removing it changes nothing,
but statisticians know what it is so they keep checking it.
🎮 Variable Selection Simulator
Let's simulate selecting variables from 100 potential marketing factors for your MMM.
See how different methods compare in speed, accuracy, and usefulness.
Running analysis...
The Professional's Cheat Sheet
✅ Use Granger When Someone Says:
"Does competitor pricing affect us at all?" → Test it
"How long before we see effects?" → Find the lag
"Who's the market leader?" → Check who moves first
❌ Don't Use Granger When Someone Asks:
"What's the ROI?" → Run incrementality test
"How should we allocate budget?" → Use MMM coefficients
"Is this campaign working?" → Look at actual performance
"Should we include TV spend?" → Of course you should
The Bottom Line
🏆 Winner: Simple Methods + Common Sense
🥉 Third Place: Granger Causality
❓ What's Second? Literally anything else
Time to learn Granger: 1 hour ⏰ Times you'll actually need it: Maybe twice in your career 🤷 Value of knowing when NOT to use it: Priceless 💎