Adstock Optimization
Find the Optimal Carryover Rate for Each Channel
What is Adstock Optimization?
Adstock optimization systematically tests multiple carryover rates to find the one that maximizes variable performance in your model.
Goal: Identify the adstock rate that produces highest T-statistic and R² improvement
Why Optimize Adstock?
The Problem with Guessing
Typical approach:
Use industry standard (TV = 70%)
Hope it's correct for your situation
May be suboptimal
Better approach:
Test rates 0% through 90%
Find empirically best rate
Data-driven, not assumption-driven
Impact of Wrong Rate
Too low (e.g., 30% when should be 70%):
Underestimates total effect
Lower coefficient
Appears less significant
Undervalues channel
Too high (e.g., 90% when should be 50%):
Overestimates persistence
Inflated long-term effect
May capture trend, not carryover
Misleading optimization
Optimal rate:
Maximizes statistical significance
Best model fit
Accurate effect measurement
How to Optimize Adstock
Method 1: Variable Testing Interface
Step 1: Select variable
Choose one media variable to optimize
Example: TV_Spend
Step 2: Select multiple adstock rates
Check: 0%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%
Or subset: 30%, 50%, 70% for quicker test
Step 3: Run test
Click "Test Variables"
System tests variable with each rate
Step 4: Review results
Each rate shown as separate row
Compare T-statistics
Compare R² increases
Compare p-values
Step 5: Select winner
Choose rate with:
Highest T-statistic
Lowest P-value
Highest R² increase
Step 6: Apply to model
Add variable with optimal rate in Model Builder
Method 2: Systematic Testing
Test incrementally:
Start with 0%, 30%, 60%, 90%
Identify range (e.g., 30-60%)
Test narrower: 30%, 40%, 50%, 60%
Find peak performance
Use optimal rate
Interpreting Optimization Results
Results Table
Example output:
TV_Spend
70%
4.2
0.001
+0.035
TV_Spend
60%
3.9
0.002
+0.032
TV_Spend
50%
3.5
0.005
+0.028
TV_Spend
80%
3.8
0.003
+0.031
TV_Spend
40%
3.1
0.008
+0.024
Winner: 70% adstock (highest T-stat, best R²)
Clear Winner Pattern
Characteristics:
One rate significantly outperforms
T-stat > 0.3 higher than next best
Clear peak in results
Decision: Use winning rate confidently
Flat Pattern
Characteristics:
Multiple rates perform similarly
T-stats within 0.2 of each other
No clear peak
Decision:
Choose middle of range
Or use business judgment
Variable may not need adstock
Monotonic Increase
Characteristics:
Higher rates always better
Even 90% shows improvement
No peak detected
Warning: May be capturing trend, not carryover
Decision:
Recheck data for trends
Consider detrending
Cap at 80% maximum
Optimization by Channel Type
TV Advertising
Typical optimal range: 60-80%
Why:
High frequency drives awareness
Brand recall persists
Long consideration periods
Test strategy: Focus on 50%, 60%, 70%, 80%
Digital Display
Typical optimal range: 30-50%
Why:
Moderate frequency
Some brand building
Retargeting extends effect
Test strategy: Focus on 20%, 30%, 40%, 50%
Paid Search
Typical optimal range: 0-20%
Why:
High intent, immediate action
Little brand building
Short consideration
Test strategy: Focus on 0%, 10%, 20%
Radio
Typical optimal range: 40-70%
Why:
Frequency-dependent
Moderate persistence
Awareness building
Test strategy: Focus on 40%, 50%, 60%, 70%
Advanced Optimization Techniques
Seasonal Optimization
Test by season:
Q4 optimal rate may differ from Q1
Holiday effects change carryover
Test separately if data allows
Product Category Optimization
Different products, different rates:
Impulse products: Lower rates
Considered purchases: Higher rates
Test by category if multiple products
Combined with Curves
Optimization order:
First optimize adstock rate
Then apply saturation curve
Don't optimize both simultaneously
Why: Separate effects for clarity
Common Optimization Scenarios
Scenario 1: New Channel Launch
Situation: Just started Radio, need optimal rate
Approach:
Wait for 8-12 weeks of data
Test broad range: 0%, 30%, 60%, 90%
Identify pattern
Refine with narrow range
Caution: Limited data means less reliable optimization
Scenario 2: Channel Performance Review
Situation: TV has been at 70% for years, is it still optimal?
Approach:
Test current rate (70%) plus neighbors (60%, 80%)
If neighbors better, expand testing
Update rate if clear improvement
Frequency: Re-optimize annually or when market changes
Scenario 3: Budget Optimization
Situation: Need ROI for each channel at optimal adstock
Approach:
Optimize adstock for each channel
Build model with optimal rates
Run decomposition for ROI
Use for budget allocation
Result: More accurate ROI, better optimization
Validation and Checks
After Optimization
Validate optimal rate:
✅ Statistical check:
T-stat significantly higher
P-value < 0.05
R² improvement meaningful
✅ Business check:
Rate aligns with channel behavior
Similar to industry benchmarks
Makes logical sense
✅ Diagnostic check:
VIF acceptable
Residuals well-behaved
No issues introduced
Red Flags
🚨 Optimal rate = 90%+
Likely capturing trend
Check for time trends in data
Consider detrending
🚨 Optimal rate = 0% for all channels
May indicate data issues
Check for weekly vs. monthly mismatch
Verify data quality
🚨 Wildly different from expectations
TV optimal at 10%
Search optimal at 80%
Investigate data quality
Best Practices
Testing Approach
Start broad:
Test 0%, 30%, 60%, 90% first
Identify general pattern
Narrow down from there
Test systematically:
Don't skip rates
Cover full range
Document results
One channel at a time:
Don't optimize all simultaneously
Focus on one variable
Validate before moving to next
Applying Results
Use optimal rate:
Apply in Model Builder when adding variable
Document choice and rationale
Include in model documentation
Retest periodically:
Quarterly or annually
When market changes
After major campaigns
Validate with business:
Share results with channel owners
Confirm makes sense
Get buy-in on rate
Troubleshooting
No clear optimal rate
All rates perform similarly
Solution:
Variable may not need adstock
Use 0% or business judgment rate
Focus optimization elsewhere
Optimization takes too long
Testing many rates for many variables
Solution:
Optimize one variable at a time
Use broader increments (0%, 30%, 60%, 90%)
Refine only top candidates
Results contradict business knowledge
Optimal rate doesn't match expectations
Investigation:
Check data quality
Review variable definition
Validate with stakeholders
May reveal new insights OR data issues
Key Takeaways
Optimize adstock by testing multiple rates systematically
Look for rate with highest T-statistic and R² increase
Typical ranges: TV 60-80%, Digital 30-50%, Search 0-20%
Test one variable at a time for clarity
Validate optimal rate against business logic
Re-optimize periodically as markets change
Use Variable Testing interface for efficient optimization
Apply optimal rates in Model Builder when adding variables
Last updated