Engines and methodology. How MMM Labs builds, compares and validates models

Six engines, one configuration, one comparable output. Here is what happens between your upload and a budget recommendation.

The six engines. Open-source and proprietary, side by side

Already running Robyn today? Keep it. MMM Labs runs it next to the others on the same data, so you can see where they agree.

MeridianGoogle · open source

Google's open-source Bayesian MMM, built for geo-level (regional) data.

Adaptive MMMProprietary

Built for messy, real-world data. It uses priors and per-channel limits to steady results where plain regression swings.

PyMC-MarketingOpen source

Bayesian MMM on PyMC, with flexible priors and adstock and saturation choices.

Constrained OLSProprietary

OLS with business rules built in. Coefficients are kept within limits you set, for models you can defend.

RobynMeta · open source

Meta's open-source MMM, using regularized regression with automated hyperparameter search.

OLSProprietary

Ordinary least squares on transformed media. It runs in seconds, so it suits quick reads and sanity checks.

Performance. A regional PyMC model in under 6 minutes

The same model took 98 minutes with PyMC-Marketing run locally on a 16-core laptop, on the same data.

3–16×
Faster PyMC regional models
1.5–2.6×
Faster PyMC national models

Every regional model reached a stable fit. In our testing.

ModelingPROJECTBig CheeseCONFIGURATIONNational - Logged
Data ReadinessRun ModelsLeaderboardContributions ManagerResults Compare
RUN TIME · ONE REGIONAL PYMC MODEL · SAME DATA
98MIN ELAPSED
Before98 min
Latest releaseunder 6 min
✓ STABLE FIT
SPEED-UP RANGE · PYMC
0×4×8×12×16×
3–16× faster · Regional models

In our testing.

Model comparison. Fit is only the first check

ModelingPROJECTBig CheeseCONFIGURATIONNational - Logged
Contributions ManagerResults Compare
FIT · R² · illustrativeHIGHER IS BETTER
#1ProprietaryR² 0.91
#2RobynR² 0.87
#3MeridianR² 0.83
#4PyMCR² 0.79
Fit alone can reward a model that memorizes the past. It is the first check, not the last.LEADER · PROPRIETARY

Honest accuracy

PyMC and Meridian fit every model twice: once on all data and once holding back recent weeks.

Advanced DiagnosticsPROJECTBig CheeseCONFIGURATIONNational - Logged
Model Fit SummaryContributionsPaid Media AnalysisDriver AnalysisParametersData
Actual vs. Fitted · Held-Back Weeks
ActualIN-SAMPLE FITILLUSTRATIVE

Pooling. Across markets and variable groups

Choose Separate, Partial or Full pooling. An equalizing factor stops your largest market from dominating the result.

Partial: each keeps its own effect, but thin markets borrow strength from their peers.

ModelingPROJECTBig CheeseCONFIGURATIONNational - Logged
Configurations LibraryModel SettingsData ReadinessRun ModelsLeaderboardContributions Manager
ENGINESEPARATEPARTIALFULL
Robyn To confirmTo confirmTo confirm
Meridian To confirmConfirmTo confirm
PyMC-Marketing To confirmTo confirmTo confirm
Adaptive MMM To confirmTo confirmTo confirm
Constrained OLS To confirmTo confirmTo confirm
OLS To confirmTo confirmTo confirm

Support per engine to be confirmed by Product before publishing.

Model forms. Linear or multiplicative. 17 response shapes.

Linear and multiplicative models

Switch any configuration to a log (multiplicative) model across all six engines. Contributions, ROI and response curves still come back in real-world units, never log units.

Business priors on both

CONFIRM: PRIORS ON LOG MODELS

Bring ROI evidence from lift tests and experiments into the model as business priors, on linear and log models, so results stay grounded in how your business works.

SHAPE FAMILIES · ILLUSTRATIVE
Exploratory AnalysisPROJECTBig CheeseCONFIGURATIONNational - Logged
SummarySeries ExplorerCorrelationMetric ConsistencyShare of SpendData Coverage
Response Shape · Saturation
SPEND →

Returns flatten as spend rises.

Placeholder for the curve drawings and engine support table from the Knowledge Center.

Contribution manager. Contribution, compared to what?

Always-on drivers like price and distribution never drop to zero, so their contribution depends on the reference point you measure against. Contribution Manager lets you rebase contributions without refitting the model, and labels every figure with the reporting setting behind it.

Advanced DiagnosticsPROJECTBig CheeseCONFIGURATIONNational - Logged
Model Fit SummaryContributionsPaid Media AnalysisDriver AnalysisParametersData
PRICE CONTRIBUTION · ILLUSTRATIVE
Price42%
TV30%
Search26%
Promo22%
REPORTING SETTING · VS. ZERO · NO REFIT

See the methodology on your own data