What we are building
A Marketing Mix Model estimates how marketing activity and other business conditions relate to an aggregate outcome over time. That outcome might be revenue, orders, qualified leads, subscriptions, or another measure the business can define consistently.
New to the method? Start by defining the outcome, inputs, workflow, outputs, and limitations before choosing the modeling framework.
This project begins with a postmortem of an earlier implementation, then documents the corrected build instead of presenting a finished model as a black box. The audit found a useful but uncomfortable distinction: exact reconciliation can prove that allocations add up without proving that the channel mix is valid. Every stage should produce an artifact that another person can inspect:
- a technical postmortem and defect log,
- a written decision question,
- a governed modeling table,
- a simple baseline,
- documented transformations and assumptions,
- time-aware validation results,
- uncertainty ranges,
- scenario outputs with clear limits.
The build sequence
Audit the previous implementation
Before rebuilding, inspect the last model as a system: source data, transformations, feature scales, model fitting, contribution calculation, validation, and published output. A bookkeeping validator can prove that channel totals reconcile without proving that the split between channels is statistically defensible.
The deeper engineering account, Building a Marketing Mix Model That Cannot Silently Lie, shows how observed floors, exact reconciliation, independent validators, cohort-maturity flags, source-boundary tests, and fail-soft checks were rebuilt—and which statistical defects still prevent a causal claim. The corrected build will preserve the useful reconciliation controls without carrying the old assumptions forward silently.
01 — Define the decision
Start with the decision the model must support. “Measure marketing” is too broad. “Estimate how reallocating the next quarterly budget across five controllable channels may change expected orders” is testable and bounded.
The decision determines the outcome, time grain, geography, channel detail, historical window, and level of uncertainty the model must expose.
02 — Create the data contract
The modeling table needs one stable row grain—often week by market—with consistent definitions across the full history. It should include the outcome, media inputs, prices or promotions, seasonality signals, distribution or availability changes, and major outside factors that could move demand.
The data contract records the source, owner, unit, aggregation rule, earliest trustworthy date, missing-value policy, and known breaks for every variable.
03 — Establish a baseline
Before adding media transformations or a complex Bayesian structure, fit a deliberately simple baseline. It should reveal whether the outcome series, seasonal pattern, structural breaks, and controls behave as expected.
If the simple model cannot be explained, added complexity will usually make the failure harder to see.
Decide What to Own and What to Borrow
Building an MMM does not require rejecting maintained frameworks. Google Meridian and Meta Robyn provide substantial modeling machinery, diagnostics, calibration paths, and standard outputs. They can accelerate a conventional build and serve as valuable challenger models.
Own the parts that define the business evidence: source contracts, outcome eligibility, CRM reconciliation, cohort maturity, offline exposure timing, control selection, experiment translation, release tests, and decision boundaries. Borrow framework machinery where its assumptions match the problem.
The custom system documented here exists because directly observed attribution must remain a hard floor, the unattributed residual must reconcile exactly, optional sources must fail softly, and incomplete cohorts and data seams must remain visible. Those requirements sit around the regression and would still need to be designed if Meridian or Robyn supplied the estimator.
Read Building a Marketing Mix Model That Cannot Silently Lie for the complete tradeoff and the case for using established frameworks as adversarial benchmarks rather than automatic answers.
04 — Model carryover and diminishing returns
Advertising can affect outcomes after the exposure period, and each additional unit of spend may produce less incremental response. MMM commonly represents those ideas with adstock and saturation transformations.
These choices are business assumptions expressed mathematically. We will show how the assumptions change channel contribution and scenario results instead of treating one transformation as universally correct.
05 — Validate through time
Randomly splitting time-series rows can leak future conditions into training. The validation plan should respect time: train on an earlier window, test on a later window, and repeat when the history permits it.
Prediction error matters, but it is not enough. We also need plausible signs and ranges, stable channel behavior, residual checks, sensitivity tests, and comparison with known experiments or major business events.
06 — Turn the model into scenarios
A useful MMM supports constrained questions: hold total budget constant, respect channel floors and ceilings, and compare a limited set of reallocations. Scenario results should include a central estimate, uncertainty, and the conditions under which the recommendation may fail.
The model informs a decision. It does not remove the need for tests, operational judgment, or measurement QA.
What this hub will publish
- the exact modeling-table specification
- a data-readiness and gap worksheet
- baseline model walkthroughs
- adstock and saturation explainers
- frequentist and Bayesian model comparisons
- time-based validation and sensitivity checks
- experiment calibration notes
- budget-scenario design
- model monitoring and refresh cadence
- honest postmortems when an approach does not hold up
The first modeling-table check
Before choosing a package or modeling framework, complete the MMM data-readiness worksheet. A model cannot recover definitions, history, or variation that the source data never contained.
Then read How to build an MMM without starting with the model, which defines the first practical milestone: a decision-ready data contract.
A necessary boundary
MMM estimates relationships from observational aggregate data. Results depend on the available variation, model structure, controls, priors or constraints, and quality of the outcome data. A precise-looking chart is not proof of causality.
We will treat experiments, holdouts, and business evidence as calibration and challenge mechanisms—not decorations added after the model is complete.