Skip to content

How To Compare AMMM With Other MMM Tooling

This page provides a comparison framework for evaluating AMMM against alternative MMM libraries.

Updated: 4 March 2026

AMMM is a Bayesian MMM library built on PyMC/PyTensor with optional JAX acceleration, maintained in the WPP/GroupM ecosystem, and designed around a stage-gated diagnostic workflow.

The main differentiator is not a single model form; it is the combination of:

  • explicit Bayesian parameterisation,
  • workflow-stage artefact governance,
  • machine-readable diagnostic outcomes used by downstream reporting.
CapabilityAMMM V2 statusNotes
Bayesian inference with posterior uncertaintyImplementedNUTS-based posterior with 94% HDI reporting.
Stage-gated workflow outputsImplementedNumbered stage folders 00_... to 80_....
Convergence gate artefactImplemented50_diagnostics/convergence_report.json (converged).
Calibration gate artefactImplemented50_diagnostics/calibration_report.json (well_calibrated).
Pareto k reliability diagnosticsImplemented50_diagnostics/pareto_k_summary.json, pareto_k.png.
Prior predictive plausibility checksImplemented10_pre_diagnostics/prior_predictive_*.
Response-curve-based optimisationImplementedSingle-period and multi-period optimisation workflows.
Ramp constraints (absolute and percentage)ImplementedParsed/validated via parse_and_validate_ramp_constraints.
Adstock-aware multi-period objectiveAvailable (advanced)Implemented, but treated as advanced/experimental in practice.
Prophet-assisted baseline optionImplementedUseful for stability, with documented leakage trade-off.
CRE/Mundlak baseline structureNot in V2Roadmap item for V3.

When comparing AMMM with another library, evaluate at least five dimensions:

  1. Diagnostic governance: are convergence, calibration, and influence diagnostics stage-gated and machine-readable?
  2. Baseline transparency: is the baseline specification explicit, auditable, and sensitivity-tested?
  3. Uncertainty treatment: are uncertainty intervals carried through decomposition and decision support?
  4. Optimisation realism: are constraints (bounds, ramps, multi-period effects) represented explicitly?
  5. Operational traceability: can stakeholders trace each business output back to diagnostic evidence?

Choose AMMM when you need:

  • a disciplined Bayesian workflow with auditable artefacts,
  • explicit handling of diagnostic failure modes,
  • practical optimisation on top of posterior-informed response curves,
  • clear separation between computational adequacy and causal claims.

Choose alternative tooling when your priority is a different ecosystem constraint (for example strict platform integration) and the diagnostic governance model remains acceptable for your risk profile.

No MMM library should be selected solely on predictive fit claims. Strong diagnostics and good fit do not, on their own, establish causal identification.