MMM dataset: 156 weeks to control before modelling.
Evidence status
What was reproduced, and what it does not prove.
The Noroa CSV, its model and Python verify 156 weeks, 15 variables, 12 holdout weeks and the usage prohibitions. They prove neither market performance nor causal validity.
- Technical specification of the Noroa data profile, specification in French (opens in a new tab)
v1, 31 August 2026 · CC BY 4.0 · specification in French · Method document written by marketing-mix.net. The figures are illustrative: they are not market measurements.
Key distinctions
Three conditions before calculating.
Common grain
Every variable must be aggregated on the same week, area and commercial perimeter. National spend joined to partial sales creates a relationship with no economic interpretation.
Stable definitions
Net revenue, margin, spend, impressions and promotions must keep the same calculation rule. An undocumented change of tool looks like a change of performance.
Usable variation
A channel that is constantly active with an almost fixed budget is hard to distinguish from the trend. Completeness does not replace independent variation.
Method
Deciding whether the data are fit for MMM
- 01Reconcile the perimeters
Check currency, taxes, returns, areas, products, weeks and accounting calendar for each source.
- 02Profile the variation
Measure minimum, maximum, coefficient of variation, weeks at zero and correlations between channels.
- 03Document the events
Add price changes, promotions, stockouts, openings, competition and tests that influence the outcome or serve for calibration.
- 04Reserve the validation
Isolate a recent period before any estimation and forbid its use in the learned transformations.
The Noroa data sheet
The synthetic dataset of the reference case makes visible 15 variables, two calibration events and 12 weeks reserved for validation.
| Week | Revenue € | Search € | Social € | Video € | Availability |
|---|---|---|---|---|---|
| 2023-01-02 | 764,152 | 32,584 | 27,284 | 19,674 | 0.976 |
| 2023-01-09 | 805,137 | 32,309 | 26,360 | 17,689 | 0.977 |
| 2023-01-16 | 821,854 | 33,547 | 27,121 | 16,408 | 0.973 |
Reproduction, training and method tests permitted. Sector benchmarking and claims of real performance prohibited.
A complete table that cannot separate two channels
Paid social and video were increased together for 31 weeks and reduced together for 27 weeks. Their spend correlation reaches 0.94.
Even with no missing value, the model can attribute the effect to one or the other depending on the priors. The individual contributions vary by 60%, while their sum stays stable at 11%.
Authorise a grouped social + video reading, or create variation through a test. Refuse a separate allocation based on these unstable coefficients.
Acceptance conditions
What must be true to act.
- 01
Every variable has a definition, unit, source, owner and transformation history.
- 02
The variation profile and the correlations between channels are examined before the model.
- 03
Price, promotions, distribution, availability and major events are covered.
- 04
A validation period is frozen before estimation.
Limits
What this analysis does not prove.
- The Noroa dataset is simulated and represents no real market.
- Statistical quality does not guarantee that every confounding variable has been observed.
- A long series does not correct a break of definition or an absence of variation.
References
Works cited.
- Gebru, T. et al. (2021), “Datasheets for Datasets”, Communications of the ACM.
- Wilkinson, M. D. et al. (2016), “The FAIR Guiding Principles”, Scientific Data.
- Jin, Y. et al. (2017), Google Research.