Simba documentation

Data Validation — How the Data Validator Checks Your Data

Simba includes an AI-powered Data Validator that checks your uploaded dataset against the requirements of a Bayesian Marketing Mix Model. It catches problems that would otherwise surface as poor model fit or unreliable results.

Why Data Validation Matters

The principle is simple: garbage in, garbage out. If your data contains errors, gaps, or inconsistencies, no statistical model — however sophisticated — can produce reliable insights. The Data Validator acts as a quality gate between your raw data and your model.

How the Data Validator Works

The Data Validator is not automatic. After uploading your data in the Warehouse configuration screen, click Start Validator Agent to run it. You can choose between Haiku (faster) or Sonnet (more thorough) as the AI model powering the analysis. Validation typically takes 3 to 10 minutes depending on dataset size.

What It Checks

The validator runs 10 specialized checks across your dataset:

1. Schema Validation

Checks that your data structure is valid:

• Date column is present and correctly formatted
• KPI (revenue/sales) column detected
• All columns have headers with consistent data types
• No structural issues (merged cells, hidden characters, encoding problems)

2. Frequency Diagnostics

Analyzes time series cadence:

• Confirms consistent weekly or daily intervals
• Detects plan-spreading patterns (e.g., even weekly splits of monthly budgets)
• Identifies gaps in the date sequence

3. Alignment

Checks that media and KPI variables align:

• Date ranges match across all variables
• No misaligned periods between spend and outcome data

4. Multiplier Validation

Validates multiplier columns for panel or hierarchical data:

• Checks multiplier column logic and consistency
• Verifies scaling relationships

5. Controls Validation

Reviews control variables:

• Appropriate variable types and variation
• Sufficient coverage across the modeling period

6. Coverage Analysis

Identifies gaps and inactive periods:

• Missing values by column and time period
• Zero-spend patterns and inactive channel periods
• Channels with insufficient variation

7. Outlier Detection

Flags statistical outliers using IQR (interquartile range — the spread between the 25th and 75th percentile of values)-based methods:

• Unusually high or low values on numeric columns
• Values that may indicate data errors vs genuine business events

8. Multicollinearity

Analyzes correlation between media channels:

• Variance Inflation Factor (VIF) analysis — VIF measures how much each channel’s spend pattern overlaps with other channels; high VIF means the model struggles to tell channels apart
• Pairwise correlation detection between channels
• Flags channels that always spend together (model cannot distinguish their effects)

9. Leakage Detection

Checks for data leakage:

• Variables that could leak future information into the model
• Columns that are derived from the target variable

10. Documentation Audit

Reviews column naming and documentation quality:

• Column naming conventions
• Descriptive header quality

Interpreting Results

The results modal displays findings in three severity levels:

Severity	Meaning
Errors	Critical issues that should be resolved before modeling — e.g., missing KPI column, severe multicollinearity, data leakage
Warnings	Issues worth investigating — e.g., high outlier count, moderate channel correlation
Info	Observations and recommendations for improvement — e.g., naming suggestions, coverage notes

Each issue includes:

• A description of the problem
• The affected columns and rows
• A specific recommendation for how to fix it
• The validation category it belongs to

When category-level insights are available, the modal also shows summary findings and action items grouped by validation category.

Common Issues and Fixes

Issue	Impact	Fix
Missing KPI column	Blocking — model cannot fit without a target	Add a revenue or sales column
High VIF between channels	High — model cannot distinguish channel effects	Consider combining correlated channels or removing one
Plan-spreading pattern	Moderate — inflates contribution estimates	Use actual weekly spend data instead of evenly split monthly budgets
Missing values in media spend	Moderate — gaps reduce model accuracy	Fill with 0 if no spend occurred, or source the missing data
Outlier in revenue	Moderate — can skew model coefficients	Verify if real (e.g., Black Friday) or a data error; consider adding event flags
Low variation in a channel	Low — model cannot estimate impact well	Consider combining with similar channels

Tips for Clean Validation Results

1. Prepare data carefully before upload — see Data Preparation
2. Use zero instead of blank for periods with no activity
3. Include at least 52 weeks of history for reliable seasonal patterns
4. Separate channels where possible rather than combining
5. Check for multicollinearity — if two channels always spend together, the model cannot distinguish their effects

Re-Running the Validator

You can re-run the Data Validator at any time after making changes to your uploaded data by clicking Start Validator Agent again. Validation results are displayed in the modal during your session and are not stored permanently.

→ Next: Supported Channels

Data Requirements

---

*See also: Data Validator Guide

Data Preparation*

This site is open source. Improve this page.