The Data Validator is an AI agent that inspects your CSV dataset against the requirements of a Bayesian Marketing Mix Model. It catches problems that would otherwise surface as poor model fit or unreliable results --- issues like missing spend data for key weeks, multicollinearity between channels, plan spreading artifacts, or data leakage.
Note: The Data Validator is not automatic. You trigger it manually from the Model Warehouse by clicking Start Validator Agent after uploading your CSV file.
Upload your CSV file (50MB max --- Excel is not supported), then click the Start Validator Agent button to open the model selection dialog.
| # | Element | Description |
|---|---|---|
| 1 | Data Source Configuration header | Step 1 of the model creation wizard --- upload your dataset here before any modeling |
| 2 | CSV upload area | Drag-and-drop or click to browse. Accepts CSV files only (.csv), maximum 50 MB. Excel (.xlsx) is not supported |
| 3 | Start Validator Agent button | Opens the model selection dialog below. Disabled until a file is uploaded. The Use Demo File button (left) loads a sample dataset for testing |
| 4 | Model selection dialog | Choose between two AI models for validation: Claude Haiku 4.5 (fast validation, 2-3 minutes, best for standard datasets) or Claude Sonnet 4.5 (deep analysis, 4-6 minutes, best for complex datasets with multiple hierarchies or unusual patterns) |
After clicking Start Validation, a progress modal shows live status updates. The full analysis typically takes 3-5 minutes.
When validation completes, a split-pane results modal displays all findings organized by severity.
| # | Element | Description |
|---|---|---|
| 1 | Summary badges | Top-level status: green "Valid Dataset" or red "Issues Found", plus counts for each severity level (Errors, Warnings, Info) |
| 2 | Issues list (left pane) | Scrollable list of all findings. Each issue shows a severity icon, message, affected columns (as code badges), and category tag. Click any issue to see details |
| 3 | Issue detail (right pane) | Selected issue expanded: severity-colored alert box, specific recommendation explaining what to fix and why it matters for MMM, plus category-level insights with key findings and action items |
| 4 | Footer actions | Close/Skip button (left) and Apply Transformations button (right, if the Data Transformer has suggestions to apply) |
| Level | Icon | Color | Meaning | Examples |
|---|---|---|---|---|
| Error | AlertCircle | Red | Critical --- prevents reliable model fitting | Missing date/KPI column, VIF > 10, duplicate columns, negative media spend |
| Warning | AlertCircle | Yellow | Quality impact --- model runs but accuracy suffers | High correlation (r > 0.9), plan spreading, sparse campaigns (>70% zeros), structural breaks |
| Info | Info | Blue | Informational --- good to know, no immediate action needed | Column inventory, data completeness %, hierarchy detection |
The validator runs 10 specialized checks in parallel, each examining a different aspect of data quality for MMM.
| # | Category | What It Checks | Key Thresholds |
|---|---|---|---|
| 1 | Schema | Date/KPI/media column detection, date parsing, duplicates, naming conventions, missing values, negative spend | Missing date or KPI = ERROR. Missing > 50% = ERROR, > 10% = WARNING |
| 2 | Frequency | Cadence detection (daily/weekly/monthly), monthly plan spreading (even weekly splits from monthly budgets), zero-variance columns | Plan spreading (cv < 0.01 in 4-week windows) = WARNING |
| 3 | Spend/Exposure Alignment | Spend ↔ exposure pairing, misalignment (spend > 0 but exposure = 0), Spearman correlation, sparse campaigns | Misaligned periods = WARNING, correlation r < 0.3 = WARNING, > 70% zeros = WARNING |
| 4 | Multiplier | Revenue vs volume KPI detection, multiplier variance (CV), zero and negative values | CV > 0.5 = WARNING, zeros = ERROR, negatives = ERROR |
| 5 | Controls | Price series (negatives, > 20% jumps), distribution range (0-1 or 0-100), promotion outliers | > 20% price jump = WARNING, negatives = ERROR, max > 5× p99 = WARNING |
| 6 | Coverage | Sample size, media coverage percentage, hierarchy column detection | < 52 rows = ERROR, < 104 rows = WARNING, < 20% media coverage = WARNING |
| 7 | Outliers | IQR-based spike detection (3× IQR bounds), structural breaks between data halves | > 5% outliers = ERROR, < 5% = WARNING, > 50% mean shift = WARNING |
| 8 | Multicollinearity | Correlation matrix (> 0.9 pairs), VIF calculation, zero-variance channels | VIF > 10 = ERROR, VIF 5-10 = WARNING, r > 0.9 = WARNING, variance = 0 = ERROR |
| 9 | Leakage & Timing | Same-period KPI-media correlation, duplicate columns | r > 0.95 same-period = WARNING, identical columns = ERROR |
| 10 | Documentation | Data completeness audit (% rows with no missing values), column inventory by type, metadata trail | Always INFO --- audit trail for reproducibility |
For detailed technical explanations of each category, see Data Validation (Technical Reference).
| Issue | Impact on Model | How to Fix |
|---|---|---|
| Missing date or KPI column | Model cannot fit at all | Ensure your CSV has a date column and a target KPI column (e.g., revenue, sales, units) |
| High multicollinearity (VIF > 10) | Coefficient uncertainty inflated, individual channel effects unidentifiable | Remove one of the correlated variables, include lift tests or strengthen the prior assumptions |
| Plan spreading (even weekly splits) | Adstock and saturation estimates will be biased because there's no real weekly variation | Request actual weekly delivery data from your media agency instead of monthly plans divided by 4 |
| Sparse campaigns (> 70% zeros) | Unstable ROI estimates due to insufficient data points | Aggregate to a higher time granularity (monthly), or exclude the channel and add back later with more data |
| Structural break in KPI (> 50% shift) | Model may not generalize across the break | Investigate the cause (business event, data error). Consider modeling periods separately or adding a structural break event |
| Negative media spend | Violates the assumption that more spend = more outcome | Fix data errors. If these are refunds/credits, net them against positive spend periods |
| Data leakage (r > 0.95 same-period) | Model "learns" from the outcome, producing inflated and meaningless ROI | Verify the column isn't derived from the KPI. Remove or lag the variable appropriately |
After reviewing validation results, you can apply automated fixes using the Data Transformer. Click Apply Transformations in the results footer to choose a transformation mode:
| Mode | Time | Scope | Best For |
|---|---|---|---|
| Fix Errors Only | 2-4 min | ~5-10 transforms | Resolving validation errors and warnings identified by the validator |
| Comprehensive (default) | 4-6 min | ~15-30 transforms | Full cleanup: error fixes plus feature engineering, normalization, and model preparation |
| Enhancements Only | 3-5 min | ~10-20 transforms | Skip error fixes, apply only feature engineering and advanced transformations |
You can also provide custom instructions (e.g., "Create 7-day rolling average for revenue", "Apply log transformation to all spend columns"). After the transformer runs, you review and select which new columns to keep --- original columns are always preserved.
- Use consistent date formats across your file. Mixed formats are the most common cause of schema errors.
- Include at least 104 weeks (2 years) of data when possible. The validator flags < 52 weeks as an error and < 104 weeks as a warning because short time series reduce seasonality detection and model reliability.
- Ensure spend variation per channel. Channels with constant spend (zero variance) cannot be identified by any model --- the validator flags these as errors.
- Pair spend with exposure metrics (e.g.,
tv_spendwithtv_impressions). The alignment validator checks for mismatched periods and low correlation between paired columns. - Break out platforms individually (e.g., Facebook and Instagram as separate columns rather than combined "Social") for more granular attribution and to reduce multicollinearity.
Validation results are displayed in the modal during your session and are not stored permanently. After making changes to your data:
- Re-upload the corrected CSV file
- Click Start Validator Agent again
- Confirm all previous errors and warnings are resolved
Iterate until you see the green Valid Dataset badge with zero errors.
Platform guides:
- Model Creation Wizard --- Continue to model setup after validation
- Model Configuration --- Configure priors and transforms
- Measurement --- Interpreting model results
Core concepts:
- Bayesian Modeling --- Why Bayesian models need clean data
- Priors & Distributions --- How validation informs prior choices
- Saturation Curves --- Why spend/exposure alignment matters
- Adstock Effects --- Why plan spreading distorts carryover estimates
Data reference:
- Data Validation (Technical) --- Deep dive on each validation category
- Data Requirements --- What data you need
- Data Preparation --- Cleaning and formatting best practices