Date Filtering

Control Which Time Periods Are Included in Your Model

Overview

Date filtering allows you to include or exclude specific observations (weeks, months) from your model without deleting data. This is essential for seasonal analysis, campaign evaluation, outlier removal, and testing model stability across different time periods.

Purpose: Selectively include/exclude time periods from modeling while preserving complete dataset

Access: Model Library → Select one model → Click "Filter Model Period"

What is Date Filtering?

The Concept

Your dataset: 104 weeks of data (2 years)

Without filtering:

• Model uses all 104 weeks

• Estimates coefficients based on full period

• R² calculated across all observations

With filtering:

• Select which weeks to include (e.g., only Q4: 26 weeks)

• Model estimates using only included weeks

• R² calculated on filtered subset

• Excluded data preserved, not deleted

Why Filter?

Seasonal analysis:

• Model Q4 separately (holiday effects different)

• Compare summer vs. winter coefficients

• Understand seasonal variation in channel effectiveness

Campaign evaluation:

• Include only campaign period

• Exclude pre-launch baseline

• Isolate campaign-specific effects

Outlier removal:

• Exclude periods with unusual events

• Remove weeks with data quality issues

• Improve model stability

Model validation:

• Train on first 80% of data

• Test on last 20%

• Assess out-of-sample performance

Data quality:

• Exclude weeks with missing values

• Remove periods before channels launched

• Clean modeling dataset

How Date Filtering Works

Observation Mask

Mechanism: Binary mask (1 or 0) for each observation

1 = Include:

• Observation used in model

• Contributes to coefficient estimation

• Included in R² calculation

0 = Exclude:

• Observation NOT used in model

• Skipped in estimation

• Not in R² calculation

• Data still exists in dataset

Example mask:

Week 1: 1 (include)
Week 2: 1 (include)
Week 3: 0 (exclude - outlier)
Week 4: 1 (include)
...

Data Preservation

Critical: Original data never deleted

What gets filtered:

• model_data (used for regression fitting)

What stays intact:

• original_data (complete dataset)

• data (full dataset for charts, decomposition)

Benefit:

• Can un-filter anytime

• Full data available for visualization

• Decomposition can use all periods

• No data loss

Using the Filter Interface

Step 1: Open Filter Dialog

In Model Library:

1. Select ONE model (checkbox)

2. Click "Filter Model Period" button

3. Filter panel opens below

Shows:

• Model name

• List of all observations

• Current inclusion status (1 or 0)

Step 2: View Observations

Table displays:

• Date/Observation: Time period identifier

• KPI Value: Dependent variable value for that period

• Included: Current status (1 or 0)

• Checkbox: Select observation for batch changes

Information:

• Total observations in dataset

• Currently included count

• Currently excluded count

Step 3: Modify Inclusion

Individual observation:

1. Click on "Included" value for specific observation

2. Toggles between 1 and 0

3. Change reflects immediately in interface

Multiple observations:

1. Check boxes for observations to change

2. Click "Include Selected" or "Exclude Selected" button

3. All selected observations change together

Select All:

1. Check "Select All" header checkbox

2. All observations selected

3. Apply Include or Exclude to all

Step 4: Apply Filter

Click "Apply Filter" button

What happens:

1. Observation mask saved to model

2. Model refits using only included observations

3. New R² calculated

4. Coefficients re-estimated

5. All statistics updated

Confirmation:

• Success message shows

• "X of Y observations included" displayed

• Model statistics reflect filtered data

Step 5: View Results

Updated model shows:

• New R² based on filtered data

• New coefficients

• Statistics from included observations only

Compare to original:

• Clone model before filtering

• Run both filtered and unfiltered versions

• Compare coefficients and R²

Common Filtering Scenarios

Scenario 1: Seasonal Model (Q4 Only)

Goal: Understand holiday-period marketing effectiveness

Process:

1. Open filter for model

2. Identify Q4 weeks (Oct-Dec)

3. Exclude all other weeks (set to 0)

4. Include only Q4 weeks (set to 1)

5. Apply filter

Result:

• Model estimates Q4-specific coefficients

• TV may have higher coefficient in Q4 (gift buying)

• Digital may be more effective (online shopping surge)

Business use:

• Optimize Q4 budget allocation

• Understand seasonal channel mix

• Plan holiday campaigns

Scenario 2: Campaign Evaluation

Goal: Measure incremental impact of 8-week campaign

Process:

1. Identify campaign weeks (e.g., Weeks 40-47)

2. Exclude pre-campaign baseline (Weeks 1-39)

3. Exclude post-campaign (Weeks 48-52)

4. Include only campaign period (Weeks 40-47)

5. Apply filter

Result:

• Model shows campaign-period effects

• Isolates campaign contribution

• Removes seasonality and trends from other periods

Business use:

• Calculate campaign ROI

• Evaluate channel effectiveness during campaign

• Justify campaign investment

Scenario 3: Outlier Removal

Goal: Exclude unusual weeks that skew results

Problem weeks:

• Week 12: Major snowstorm, sales abnormally low

• Week 35: System outage, data incomplete

• Week 52: Holiday closure, partial week

Process:

1. Identify problem weeks

2. Set those weeks to 0 (exclude)

3. Keep all other weeks at 1 (include)

4. Apply filter

Result:

• More stable coefficient estimates

• Higher R² (outliers removed)

• Model represents "normal" periods

Business use:

• Cleaner model for forecasting

• Avoid outlier-driven insights

• Robust optimization recommendations

Scenario 4: Train/Test Split

Goal: Validate model on hold-out data

Process:

1. Keep first 80% as training (Weeks 1-83)

2. Exclude last 20% as test set (Weeks 84-104)

3. Fit model on training data only

4. Later use model to predict test set

Result:

• Model built without seeing test data

• Can assess out-of-sample performance

• Realistic validation of model quality

Business use:

• Confidence in model predictions

• Avoid overfitting

• Demonstrate forecast accuracy

Best Practices

Before Filtering

Document baseline:

• Note original R² and coefficients

• Export model before filtering

• Keep record of unfiltered version

Validate filtering rationale:

• Clear business reason for filtering

• Documented outlier identification

• Statistical or domain justification

During Filtering

Minimum observations:

• Keep at least 26 observations (rule of thumb)

• More observations = more reliable estimates

• Need ~10 observations per variable minimum

Be conservative:

• Don't exclude too much data

• Every excluded observation reduces power

• Balance data quality vs. sample size

Visual check:

• Review which observations excluded

• Ensure makes sense

• Look for unintended patterns

After Filtering

Compare results:

• Filtered vs. unfiltered coefficients

• How much did they change?

• Do changes make sense?

Check R²:

• R² increased: Good (removed noise/outliers)

• R² similar: Filter had little impact

• R² decreased: May have removed signal

Document changes:

• Which observations excluded

• Why they were excluded

• Impact on coefficients and R²

Re-including Data

To undo filter:

1. Open filter interface

2. Set all observations to 1 (include)

3. Apply filter

4. Model returns to using full dataset

Effect:

• Coefficients revert to original estimates (approximately)

• R² returns to full-data value

• Model back to baseline state

Advanced Filtering Techniques

Rolling Window Analysis

Goal: Test model stability over time

Process:

1. Filter to first 52 weeks, run model

2. Filter to weeks 13-64, run model

3. Filter to weeks 25-76, run model

4. Compare coefficients across windows

Insight:

• Are coefficients stable over time?

• Do channel effects change?

• Is model robust or time-dependent?

Seasonal Comparison

Create separate models:

• Model A: Q1 only

• Model B: Q2 only

• Model C: Q3 only

• Model D: Q4 only

Compare:

• Which channels perform best in each quarter?

• How do coefficients vary seasonally?

• Should you model seasons separately?

Pre/Post Analysis

Filter for event study:

• Before change (e.g., new pricing strategy)

• After change

• Compare coefficients

Measures:

• Did marketing effectiveness change?

• Did channel mix optimal change?

• What's the impact of structural change?

Troubleshooting

"Need at least 26 observations"

Cause: Excluded too many observations

Solution:

• Include more observations

• Minimum 26 needed for reliable estimation

• Ideally 52+ (one year of weekly data)

Model fit deteriorated after filtering

Cause: Removed important signal or created bias

Possible reasons:

• Removed periods with high variance (needed for estimation)

• Created artificial patterns

• Reduced sample diversity

Solution:

• Review which observations excluded

• Consider more inclusive filtering

• Compare to unfiltered model

• May need to keep problematic observations

Coefficients changed dramatically

Cause: Filtered periods had different relationships

Examples:

• Excluded holiday periods where TV is more effective

• Removed summer when Digital dominates

• Filtered out campaign periods with different dynamics

Solution:

• This may be legitimate (showing time-varying effects)

• Or may indicate filtering was too aggressive

• Consider if changes make business sense

• Document and explain to stakeholders

Can't un-filter (all observations still excluded)

Cause: Filter dialog not updating

Solution:

• Refresh page

• Reopen filter interface

• Select all observations

• Set to 1 (include)

• Apply filter again

Filtering vs. Date Range Selection

Two Different Approaches

Date Range (simple):

• Include: 2024-01-01 to 2024-06-30

• Exclude: Everything outside this range

• Continuous period

• Faster to configure

Observation Mask (flexible):

• Include: Weeks 1, 3, 5, 7 (custom selection)

• Exclude: Weeks 2, 4, 6, 8

• Non-continuous periods allowed

• Fine-grained control

When to Use Each

Use Date Range when:

• Simple continuous period needed

• Seasonal split (Q1, Q2, etc.)

• Train/test split

• Campaign period analysis

Use Observation Mask when:

• Excluding specific outlier weeks

• Non-continuous periods

• Complex patterns (every other week)

• Removing scattered bad-data observations

Note: MixModeler primarily uses observation mask for flexibility. Date range filtering is a special case that can be converted to mask.

Integration with Other Features

With Model Comparison

Compare filtered vs. unfiltered:

1. Clone model

2. Filter clone to subset

3. Use Model Comparison to see differences

4. Understand impact of filtering

Business insight: Does relationship hold across all periods or only subset?

With Decomposition

Decomposition behavior:

• Uses FULL dataset for decomposition

• Even if model filtered for estimation

• Allows visualizing contribution over all periods

Why:

• Stakeholders want to see complete picture

• Filtered periods still get predictions

• Can assess model performance on excluded data

With Diagnostics

Diagnostics run on:

• Only INCLUDED observations

• Tests based on filtered dataset

• R² and statistics reflect filtered model

Consideration:

• Fewer observations → less powerful tests

• May affect diagnostic conclusions

• Keep sufficient sample for reliable diagnostics

Export and Reimport

Filtering in Excel Export

What's exported:

• Observation Mask sheet with inclusion status

• Complete dataset (all observations)

• Model fitted on filtered subset

Excel contains:

• Column "Included_In_Model" (1 or 0)

• All 104 observations listed

• Shows which were used for estimation

Reimporting Filtered Models

On reimport:

• MixModeler reads Observation Mask sheet

• Recreates filter exactly

• Fits model using same filtered subset

• Perfect replication of filtered model

Benefits:

• Share filtered models with colleagues

• Archive specific analyses

• Document filtering decisions permanently

Key Takeaways

• Date filtering includes/excludes observations without deleting data

• Each observation has 1 (include) or 0 (exclude) status

• Minimum 26 observations recommended for reliable estimation

• Model refits automatically when filter applied

• Original data always preserved, can un-filter anytime

• Use for seasonal analysis, campaign evaluation, outlier removal

• Compare filtered vs. unfiltered models to understand impact

• Decomposition uses full dataset even if model filtered for estimation

• Export includes filter settings for perfect reimport

• Document filtering rationale and impact for transparency