# Performance Optimization

### Overview

Optimizing your MixModeler workflow saves time and enables faster iteration. This guide covers strategies to maximize efficiency from data preparation through final insights delivery.

### Hardware Optimization

#### Recommended Setup

**For Best Performance**:

* **RAM**: 16 GB minimum, 32 GB recommended
* **CPU**: Modern quad-core (3.0+ GHz)
* **GPU**: Dedicated GPU (NVIDIA/AMD) for 10-100x speedup
* **Browser**: Chrome or Edge (latest version)
* **Internet**: 5+ Mbps for initial load (then works offline)

**Budget Option** (Still Works):

* **RAM**: 8 GB
* **CPU**: Dual-core (2.5+ GHz)
* **GPU**: Integrated graphics (WASM-only acceleration)
* **Browser**: Any modern browser
* **Internet**: 1+ Mbps

#### GPU Acceleration

**Enable GPU Processing**:

1. Use Chrome or Edge (best GPU support)
2. Update graphics drivers
3. Enable hardware acceleration in browser settings
4. Verify GPU badge appears in MixModeler toolbar

**Expected Speedups**:

* Small models (30 vars): 5-10x faster
* Medium models (100 vars): 15-30x faster
* Large models (300+ vars): 30-100x faster

**If No GPU**:

* WASM acceleration still provides 5-10x speedup
* All features work, just slower
* Consider cloud GPU instance for large models

#### Browser Settings

**Chrome/Edge Optimization**:

1. **Enable Hardware Acceleration**:
   * Settings → System → Use hardware acceleration
   * Toggle ON
   * Restart browser
2. **Allow Sufficient Memory**:
   * Close unnecessary tabs
   * Disable memory-intensive extensions
   * Restart browser periodically
3. **Enable WebGPU** (if available):
   * `chrome://flags` → Search "WebGPU"
   * Enable "Unsafe WebGPU"
   * Restart browser

### Data Preparation Optimization

#### Minimize Pre-Processing

**In Excel**:

* Keep one clean sheet (delete others)
* Remove unnecessary columns before upload
* Pre-calculate any complex formulas
* Save as `.xlsx` (not `.xls` or `.xlsm`)

**Benefits**:

* Faster upload
* Reduced file size
* Cleaner data validation

#### Optimal Data Size

**Sweet Spot**:

* Variables: 20-100 (most projects)
* Observations: 52-156 weeks
* File size: <5 MB

**If Larger Dataset**:

* Remove unused variables
* Aggregate to weekly (if daily)
* Focus on relevant time period
* Split into multiple projects if needed

**Performance Impact**:

| Dataset Size       | Upload Time | Model Fitting |
| ------------------ | ----------- | ------------- |
| 30 vars × 52 obs   | <1 sec      | 0.5-1 sec     |
| 100 vars × 104 obs | 1-2 sec     | 1-3 sec       |
| 300 vars × 260 obs | 3-5 sec     | 5-15 sec      |

### Workflow Efficiency

#### Start Simple, Scale Up

**Phase 1: Quick Baseline** (30 minutes)

1. Upload data
2. Select top 5-10 variables
3. Build simple OLS model
4. Quick diagnostic check

**Phase 2: Refinement** (2-3 hours)

1. Add control variables
2. Test transformations
3. Run full diagnostics
4. Compare 3-5 model specifications

**Phase 3: Finalization** (1-2 hours)

1. Optimize best model
2. Generate decomposition
3. Create reports
4. Prepare presentation

**Benefits**:

* Early feedback on data quality
* Catch issues before deep work
* Iterative improvement
* Faster to useful insights

#### Use OLS Before Bayesian

**Speed Comparison**:

* OLS model: 0.5-3 seconds
* Bayesian model: 60-600 seconds (2-10 minutes)

**Strategy**:

1. Build and refine with OLS (fast iteration)
2. Run diagnostics on OLS version
3. Optimize transformations with OLS
4. Only switch to Bayesian when OLS model finalized
5. Run Bayesian once for final uncertainty quantification

**Time Saved**: Hours during model development

#### Batch Operations

**Variable Workshop**:

* Select multiple variables (checkbox)
* Apply same transformation to all
* Set group assignments in bulk
* Faster than one-by-one

**Variable Testing**:

* Test multiple adstock rates simultaneously
* Run all diagnostic tests at once
* Compare multiple models side-by-side

**Model Library**:

* Clone models instead of rebuilding
* Export multiple models at once
* Batch delete old iterations

### Model Building Efficiency

#### Reuse and Clone

**Clone Existing Models**:

1. Find similar past model in library
2. Click "Clone"
3. Modify variables/settings
4. Fit new model

**Benefits**:

* Preserves transformations
* Copies decomposition groups
* Saves variable selection time
* Maintains best practices

**Use Cases**:

* Test different time periods
* Add/remove few variables
* Compare OLS vs Bayesian
* Try different adstock rates

#### Variable Testing First

**Before Adding to Model**:

1. Go to Variable Testing
2. Test variable significance
3. Optimize adstock rate
4. Check VIF with existing variables

**Add to Model Only If**:

* Significant in testing
* Low VIF (<5)
* Makes business sense
* Improves model fit

**Saves Time**:

* Avoid adding useless variables
* Reduce model iterations
* Faster to final model

#### Incremental Building

**One Variable at a Time**:

* Add variable
* Check impact on R², coefficients
* Run quick diagnostics
* Keep or discard
* Document decision

**Faster Than**:

* Adding 10 variables at once
* Debugging which ones caused problems
* Removing one-by-one to find culprit

#### Diagnostic Shortcuts

**Quick Check** (30 seconds):

* Run only VIF test
* Check R² and Adjusted R²
* Review coefficient signs

**Full Validation** (2-3 minutes):

* Run all diagnostic tests
* Only on serious model candidates
* Not every iteration

**When to Skip Diagnostics**:

* Very exploratory models
* Testing single variable impact
* Preliminary comparisons

**Always Run Before**:

* Final model selection
* Presenting to stakeholders
* Making business decisions

### Bayesian Model Optimization

#### Fast Inference Mode

**For Rapid Exploration**:

* Enable "Fast Inference" toggle
* Uses SVI instead of MCMC
* 10-20x faster than standard Bayesian
* Approximate but good for iteration

**Workflow**:

1. Explore with Fast Inference
2. Compare multiple specifications
3. Select best model
4. Run full MCMC on final choice only

**Time Saved**: 80-90% during exploration

#### MCMC Settings Optimization

**For Quick Testing**:

* Chains: 2 (instead of 4)
* Draws: 1,000 (instead of 2,000)
* Tune: 500 (instead of 1,000)

**Runtime**: 50-60% faster

**For Final Model**:

* Chains: 4
* Draws: 2,000-3,000
* Tune: 1,000-1,500

**Check Convergence**: Always verify R-hat <1.01

#### Prior Optimization

**Start with Defaults**:

* Weakly informative priors
* Let data drive estimates
* Fast convergence

**Use Informative Priors When**:

* Small dataset (n <75)
* Convergence issues
* Strong prior knowledge

**Avoid**:

* Over-specifying priors unnecessarily
* Testing many prior configurations
* Use priors strategically, not exhaustively

### Large Dataset Strategies

#### Variable Reduction

**If 200+ Variables**:

**Strategy 1: Pre-screen**

* Calculate correlation with KPI
* Keep top 50 by correlation
* Add back theoretically important variables

**Strategy 2: Group and Aggregate**

* Combine similar channels
* Create category totals
* Reduce from 200 to 30-50 key variables

**Strategy 3: Staged Modeling**

* Model by category (Digital, TV, Print separately)
* Identify top performers in each
* Build combined model with winners

**Benefits**: 5-10x faster processing

#### Chunking Time Periods

**For Very Long Time Series** (300+ weeks):

**Approach 1: Rolling Window**

* Model most recent 104 weeks
* Update quarterly with new data
* Drop oldest data

**Approach 2: Split Sample**

* Build separate models per year
* Compare coefficients over time
* Identify structural changes

**Approach 3: Monthly Aggregation**

* Aggregate weekly to monthly
* Reduces observations by 75%
* Loses some granularity but much faster

#### Memory Management

**If Running Out of Memory**:

1. **Close Other Applications**
   * Especially browser tabs
   * Memory-intensive programs
   * Background processes
2. **Restart Browser**
   * Clears memory leaks
   * Fresh start
   * Do every few hours of heavy use
3. **Reduce Model Size**
   * Fewer variables
   * Shorter time period
   * Simplify transformations
4. **Use Faster Hardware**
   * More RAM (16-32 GB)
   * Faster CPU
   * SSD instead of HDD

### Reporting Optimization

#### Template Creation

**Build Once, Reuse**:

**Excel Template**:

1. Export best model
2. Add executive summary sheet
3. Format beautifully
4. Save as template
5. Reuse for future models (replace data)

**PowerPoint Template**:

1. Create slide deck structure
2. Add placeholder charts
3. Save as template
4. Update with new model results

**Benefits**: 70-80% less time on reporting

#### Automated Exports

**Batch Export**:

1. Export model to Excel (includes all data)
2. Generate decomposition (included in export)
3. Create PDF diagnostics for key tests
4. Package together for stakeholders

**One-Click Package**:

* Model specifications
* Diagnostic validation
* Attribution insights
* Ready to share

#### Visualization Reuse

**Save Chart Configurations**:

* Screenshot favorite views
* Document chart settings
* Recreate quickly in new models

**Export Charts**:

* Right-click → Save Image
* Insert into presentations
* No need to recreate externally

### Collaboration Efficiency

#### Model Sharing

**Export-Import Workflow**:

**Analyst A**:

1. Build model
2. Export to Excel
3. Share file with Analyst B

**Analyst B**:

1. Upload same data
2. Import model from Excel
3. Modify and iterate
4. Export and share back

**Use Cases**:

* Team collaboration
* Peer review
* Client deliverables
* Version control

#### Documentation Standards

**Consistent Naming**:

```
ProjectName_ModelType_Version_Date.xlsx

Examples:
Q4_Campaign_OLS_v1_20250104.xlsx
Annual_Bayesian_Final_20250115.xlsx
```

**Benefits**:

* Easy to find files
* Clear version history
* Team alignment

**Model Notes**:

* Include README in exports
* Document key decisions
* List any data issues
* Note model limitations

### Performance Monitoring

#### Track Your Speed

**Benchmark Standard Operations**:

| Task                   | Target Time | Your Time |
| ---------------------- | ----------- | --------- |
| Data upload (100 vars) | <3 sec      | \_\_\_    |
| OLS model fit          | <2 sec      | \_\_\_    |
| Full diagnostics       | <5 sec      | \_\_\_    |
| Bayesian MCMC          | <5 min      | \_\_\_    |
| Decomposition          | <3 sec      | \_\_\_    |

**If Slower Than Target**:

* Check GPU acceleration active
* Close other applications
* Restart browser
* Consider hardware upgrade

#### Identify Bottlenecks

**Common Slowdowns**:

1. **Data Upload** (>5 sec):
   * File too large
   * Too many unnecessary variables
   * Network issues
2. **Model Fitting** (>10 sec OLS):
   * Too many variables
   * Large dataset
   * No acceleration active
3. **Diagnostics** (>10 sec):
   * Complex tests
   * Large models
   * CPU bottleneck
4. **Bayesian** (>15 min):
   * Too many draws
   * Convergence issues
   * Settings too conservative

**Solutions**: See respective sections above

### Workflow Automation

#### Keyboard Shortcuts

**Learn Common Actions**:

* Ctrl/Cmd + S: Save model
* Ctrl/Cmd + C/V: Copy/paste
* Click + Drag: Select multiple variables
* Shift + Click: Select range

**Saves**: Seconds per action, hours over project

#### Browser Bookmarks

**Bookmark Key Pages**:

* Data Upload
* Model Builder
* Model Library
* Variable Workshop

**Quick Navigation**: Click bookmark instead of menu navigation

#### Workspace Organization

**Single Monitor**:

* Use browser tabs efficiently
* Keep MixModeler in dedicated window
* Excel/PowerPoint in separate window

**Dual Monitor**:

* MixModeler on primary monitor
* Excel/analysis on secondary monitor
* Drag-and-drop between screens

### Best Practices Summary

**Hardware**:

* Use GPU acceleration when available
* 16+ GB RAM for large models
* Keep browser updated
* Restart periodically

**Workflow**:

* Start simple, add complexity
* Use OLS before Bayesian
* Clone models to save time
* Test variables before adding

**Data**:

* Clean data thoroughly upfront
* Minimize file size
* Remove unnecessary variables
* Use optimal granularity

**Modeling**:

* Incremental building (one variable at a time)
* Quick diagnostics for iterations
* Full validation for finals
* Fast Inference for Bayesian exploration

**Reporting**:

* Create templates
* Batch export
* Reuse visualizations
* Document consistently

**Time Allocation** (for typical project):

* Data prep: 30% of time
* Model building: 40% of time
* Validation: 15% of time
* Reporting: 15% of time

**Optimize Each Phase**: Following these practices can reduce total project time by 40-60%

***

**Congratulations!** You've completed the Best Practices section. Apply these techniques to build better models faster and deliver more impactful insights to stakeholders.