Marketing Mix Modeling (MMM)

What marketing mix modeling is

Marketing mix modeling (MMM), also called media mix modeling, estimates how much each marketing channel contributes to sales, using aggregate data over time: weekly or monthly sales, spend on each channel, and other factors such as price, promotions and seasonality. Because it needs no user-level tracking or cookies, MMM has come back into fashion as privacy rules and browser changes weaken digital attribution. Open-source frameworks such as Meta’s Robyn, Google’s Meridian and LightweightMMM have made it popular well beyond large consumer-goods companies.

The core of every MMM is a regression of sales on transformed media spend. Two transformations capture how advertising really works:

• Adstock (carry-over): advertising keeps working after it is shown. A TV burst this week still drives some sales next week and the week after. Geometric adstock models this as a_t = x_t + λ a_t−1, where λ is the decay rate. With λ = 0.6, 60% of this week’s effect carries over to next week; the half-life is ln(0.5) / ln(λ) periods.
• Saturation (diminishing returns): the tenth thousand spent in a week works less hard than the first. This tool uses a logarithmic response, m·ln(1 + a/m) with m the average adstocked spend, which is close to linear at low spend and flattens as spend grows.

How this tool fits the model

1. Base terms: an intercept, an optional linear trend, optional yearly seasonality (two pairs of sine and cosine terms for weekly, monthly or quarterly data) and any control variables you select, such as price, promotions, holidays or distribution.
2. Adstock decay: either a fixed value for all channels, or estimated for each channel by searching a grid of decay rates (0 to 0.95) and keeping the combination with the smallest sum of squared errors, refined in coordinate-descent passes.
3. Regression: ordinary least squares on the transformed variables, with Newey-West standard errors that are robust to the autocorrelation typical of time-series data.
4. Sign constraint: negative media effects are almost always the result of confounding (for example spend that rises when sales fall), so by default channels with a negative estimate are dropped and the model is refitted.
5. Decomposition: each channel’s contribution in each period is its coefficient times its transformed spend; base sales are the rest of the fitted value.

The key outputs

• Contribution and share: total sales attributed to each channel over the period and its share of total sales and of the media effect.
• ROI: contribution ÷ spend, the average return per unit spent. An ROI of 2.3 means each unit of spend generated 2.3 units of sales (before costs and margins).
• Marginal ROI: the extra sales from 1% more spend divided by that extra spend. With diminishing returns, marginal ROI is lower than average ROI, and it is the right number for deciding where the next unit of budget should go.
• Response curves: total contribution of each channel as spend is scaled from 0% to 200% of the current level, showing where saturation sets in.
• Budget reallocation: a simple optimiser that moves budget from the channel with the lowest marginal ROI to the one with the highest, keeping the total budget fixed and each channel within ±50% of its current spend to avoid extrapolating far beyond the data.

Worked example with the built-in data

The example covers 104 weeks of sales with spend on TV (in flights), search and social, plus average price and yearly seasonality. The model explains 85% of the variation in weekly sales (MAPE 3.5%, Durbin-Watson 2.2). Media drive about 24% of sales. Search has the highest return (ROI 4.9, marginal ROI 3.6), TV and social have ROIs of about 2.3 and 2.0. TV has the longest carry-over (decay 0.70, half-life of about two weeks), as expected for a brand-building medium. The optimiser suggests shifting budget from TV and social towards search, raising media-driven sales by about 9% with the same total budget; this is a hypothesis to test with a geo experiment before acting on it.

Data requirements

• One row per week (preferred) or month, in time order, without gaps.
• At least two years of weekly data (100+ observations) is recommended so that seasonality and carry-over can be separated from media effects; fewer periods allow only a few channels.
• Spend should vary: channels that are always on at the same level, or always move together, cannot be separated statistically. Flighting and natural variation help.
• Include the big non-media drivers: price, promotions, distribution, competitor activity, holidays, macro shocks such as lockdowns. Omitting them biases media effects.
• Use spend for all channels in the same currency; impressions or GRPs can be used instead, but ROI is then per impression.

Validating an MMM

A good in-sample fit is necessary but not sufficient. Check that coefficients have plausible signs and magnitudes, that decay rates make sense (TV usually decays more slowly than search), and that residuals show no strong pattern. Refit the model on the first 80% of the data and compare predictions with the last 20%. Most importantly, calibrate with experiments: geo-holdout or lift tests for at least the largest channels are the best check that the model’s causal estimates are right.

Limitations of this lite MMM

This is a transparent, single-equation least-squares model. It does not include Bayesian priors, hierarchical structures across regions, interaction effects between channels, or long-term brand effects, and its uncertainty estimates do not reflect the search over decay rates. It is ideal for learning MMM, for quick first models, for academic projects and for checking the plausibility of vendor results; for high-stakes budget decisions, use it as a starting point together with experiments and fuller frameworks.

Reporting

Report the data period and frequency, the transformations (adstock type and decay, saturation), the controls, model fit (R², MAPE), and for each channel its decay, contribution, share and ROI, with a note that estimates come from observational data. Example: “A marketing mix model of 104 weeks of sales (R² = .85, MAPE = 3.5%) with geometric adstock and logarithmic saturation attributed 24.3% of sales to media. Search had the highest return on investment (4.94), followed by TV (2.28) and social (2.04).”

Glossary of MMM terms

• Base sales: sales that would occur without media, driven by trend, seasonality, price and distribution.
• Incremental sales: sales caused by a channel, its contribution in the decomposition.
• Adstock decay (λ): the share of a period’s advertising effect that carries over to the next period.
• Half-life: the number of periods for the effect of a burst to fall to half, ln(0.5)/ln(λ).
• Saturation: the flattening of the response curve as spend increases.
• Response curve: total contribution of a channel as a function of its spend.
• mROI: marginal return on investment, the return on the next unit of spend.
• Calibration: adjusting or validating model estimates with experimental lift measurements.

Key references: Jin, Wang, Sun, Chan and Koehler (2017), “Bayesian methods for media mix modeling with carryover and shape effects”, Google; Broadbent (1979) on adstock; Hanssens, Parsons and Schultz (2001), Market Response Models.