Reading the Report

An MMM Framework report is a decision document, not a diagnostics dump. This guide walks the report section by section: what each chart shows, the decision it supports, and the red flags that mean "don't act yet." Read it alongside the live example report. For the plain-language version aimed at planners and CMOs, see Interpreting Results.

Executive Summary

• Shows: the headline — blended ROI, the top and bottom channels, total modeled marketing contribution, and whether the fit is trustworthy.
• Decides: where the conversation starts. It orients the room before the detail.
• Red flag: a banner noting an approximate (MAP/ADVI) fit or non-convergence. If you see one, treat every number as provisional and re-fit with full NUTS before deciding anything.

Model Fit & posterior-predictive checks

• Shows: predicted vs actual sales over time with a predictive band, plus posterior-predictive checks — replicated datasets overlaid on the observed one, an interval-calibration curve, and posterior-predictive p-values for summary statistics.
• Decides: whether the model describes this business well enough to reason about at all. Good coverage (the actuals sit inside the band about as often as the band claims) is the license to read the rest.
• Red flag: the band systematically misses turning points, coverage is far from nominal, or a p-value is extreme (near 0 or 1) — the model is missing structure (a driver, a level shift, the wrong likelihood).

Channel ROI Estimates (the forest plot)

• Shows: each channel's return per dollar (or efficiency per unit for impression-measured channels) as a point with a credible interval — a "forest" of intervals you can rank at a glance.
• Decides: which channels are earning and which aren't. The ordering is usually more robust than the magnitudes; fund up the channels whose interval sits clearly above the break-even reference and scrutinize those below it.
• Red flag: a very wide interval (the model can't tell — often a low-variation or collinear channel) or an interval straddling break-even. Neither is a verdict; both are a call for an experiment.

Revenue Decomposition (waterfall & time series)

• Shows: how observed sales split into baseline (intercept + trend + seasonality), each media channel, and controls — as a waterfall of totals and as stacked contributions over time.
• Decides: how much of the business marketing actually moves. A large baseline and small media slice is the common, honest picture — it bounds how much upside any reallocation can capture.
• Red flag: a channel's contribution that dwarfs its spend share with no experimental support, or a baseline so small it implies marketing drives nearly all sales — usually a confounder leaking into a coefficient.

Saturation & Diminishing Returns

• Shows: each channel's response curve — incremental sales as a function of spend — with an uncertainty band, and where current spend sits on it.
• Decides: headroom. A channel on the steep part of its curve rewards more budget; one on the flat part is saturated and the next dollar buys little. This is where "should we spend more here?" gets answered.
• Red flag: current spend far past the elbow (you're paying for saturated inventory) or a curve so uncertain it can't distinguish steep from flat — the data hasn't varied spend enough to learn the shape.

Carryover Effects (Adstock)

• Shows: the posterior carryover kernel per channel — how a week's exposure decays over subsequent weeks — and the implied half-life.
• Decides: flighting and measurement timing. A long half-life (brand channels) means effects persist and an experiment needs a longer read window; a short one (performance channels) means near-immediate response.
• Red flag: a half-life implausibly long for the channel, which can trade off with the trend and inflate carryover — cross-check against the sensitivity section.

Prior vs. Posterior

• Shows: each key parameter's prior (what you assumed before data) overlaid on its posterior (what the data taught you).
• Decides: how much to trust each estimate. A posterior much narrower than its prior means the data was informative; a posterior that just reprints the prior means the data said little and the number is riding on your assumption.
• Red flag: a posterior pinned against a prior boundary, or an ROI conclusion that only holds because the prior was tight — flagged so you can't mistake an assumption for a finding.

Budget Reallocation Simulator

• Shows: the modeled outcome of shifting budget between channels, propagating saturation and uncertainty into the projected lift — with break-even spend zones from marginal ROI.
• Decides: the actual plan. Move budget toward channels with high marginal return until their curves flatten to meet the others — not toward the highest average ROI, which ignores diminishing returns.
• Red flag: a projected lift whose uncertainty band includes "no change." Reallocation this model can't distinguish from noise is a hypothesis to test, not a decision to ship.

Sensitivity Analysis

• Shows: how the headline conclusions move when the model is perturbed — dropping windows, excluding high-spend weeks, changing specification — as a spec-curve of estimates.
• Decides: how much to lean on a finding. A conclusion stable across perturbations is one you can act on; one that flips when you drop a quarter is fragile.
• Red flag: wide swings under mild perturbation. The single reported number is then a point on a cloud, and the report says so on purpose.

Recommendations & Causal Assumptions

• Shows: the concrete budget moves the analysis supports, and — crucially — the causal assumptions they rest on: which confounders were controlled, what's assumed exogenous, and where an experiment is needed to confirm.
• Decides: what to do and what to verify next. The assumptions section is the honest fine print — an MMM is a causal claim, and this is where the claim's conditions are stated.
• Red flag: a recommendation presented without its assumptions, or one that depends on a channel whose ROI interval straddles break-even. Those go to the experiment queue, not the budget.