Econometric modeling is the use of statistical models, mostly regression, to measure economic relationships from data, and in marketing it means estimating how much sales respond to advertising, price, promotions and everything else that changes week to week. In the UK, “marketing econometrics” is what much of the industry calls marketing mix modeling: The same weekly regression of sales on media, price and seasonality, under a name that comes from economics departments rather than marketing ones.
I’ve built marketing mix models for UK companies including Monzo, Time Out and The AA, and in that world the person who builds the model is as likely to be called an econometrician as a data scientist. The difference isn’t just a name. Econometrics comes with a habit of worrying about the properties of time series data, and about whether a relationship you’ve found is real or an accident of two things trending at the same time, which a lot of newer MMM work skips. I’ll show you why trends fool regressions in marketing data, with a demo you can break yourself.
Where marketing econometrics comes from
The UK became a center for marketing econometrics partly because of how its advertising industry chose to argue for itself. The Institute of Practitioners in Advertising (IPA) set up its Effectiveness Awards in 1980 to show that advertising effects could be measured, and over the decades the entries have built a databank of more than 1,500 case studies (IPA). To win, an entrant has to show that the campaign paid back, which in practice means separating the campaign’s effect from price changes, distribution, competitors and the weather. That kind of evidence is exactly what a regression on weekly sales data is built to provide, so in my view it’s no surprise that econometric modeling became a standard part of the UK agency and consultancy toolkit.
Some of MMM’s core ideas came out of that world. Simon Broadbent, then at the agency Leo Burnett in London, is usually credited with introducing “adstock” in a 1979 paper to describe how the effect of TV advertising builds and decays over the weeks after it airs (the adstock guide covers the maths). Later, Les Binet and Peter Field used the IPA Databank for their work on long-term brand building versus short-term activation, which you’ll find summarized in The Long and the Short of It. The history of marketing mix modeling traces the US side of the story, which runs through packaged goods companies and syndicated scanner data, but the methods ended up in the same place.
What econometric modeling involves in practice
At its core, a marketing econometric model is a time series regression: Weekly sales are the outcome, and the inputs are media spend (or impressions, or TV ratings) for each channel, price, promotions, distribution, seasonality, holidays and anything else that plausibly moves sales. The model estimates a coefficient for each, and those coefficients become the contribution chart, the ROI table and the response curves. Written down, a simple version looks like this:
Most of the actual work, though, isn’t in fitting that equation. It’s in checking that the equation deserves to be believed. Econometricians spend a lot of their time on four things that are easy to skip if you come at MMM from machine learning: whether the series are stationary, whether the residuals are autocorrelated, how the variables should be transformed (logs, adstock, saturation), and whether each coefficient can be identified at all, given how much the inputs moved together. I’ll take them in turn.
Stationarity: does the series have a stable level?
A stationary series fluctuates around a stable mean with a stable amount of variation, like weekly temperature anomalies or the week-to-week change in a sales series. A non-stationary series trends or wanders, so its mean depends on when you look: Revenue for a growing company, cumulative brand awareness, or a stock price. Most marketing data in levels is non-stationary in some way, because businesses grow, prices inflate and budgets ratchet up year after year.
That matters because the standard regression tests assume the errors are independent draws around a stable relationship. When both the outcome and an input are trending, the regression can attribute the shared trend to the input, and the t-statistics will tell you it’s highly significant. Econometricians test for this with unit-root tests (the Dickey-Fuller test, from 1979, is the classic), and deal with it by modeling changes instead of levels, by including an explicit trend, or, when two series really do move together over the long run, by modeling that long-run relationship directly through cointegration, which Robert Engle and Clive Granger set out in 1987 (Econometrica).
Autocorrelation: are this week’s errors related to last week’s?
Even with stationary series, the residuals of a weekly model are often autocorrelated: If the model under-predicts sales this week, it probably under-predicted last week too, because something it’s missing (a competitor’s promotion, a slow-building trend, a heatwave) persists over several weeks. The Durbin-Watson statistic is the standard first check; it sits around 2 when residuals are independent and falls towards zero when they’re positively autocorrelated. Low values mean the standard errors are too small and something is missing from the model, which is why reading the residuals is the first thing I look at after a fit, before any ROI number.
Elasticities and log models
Econometricians tend to think in elasticities (the percentage change in sales for a 1 percent change in an input) rather than in “dollars of revenue per dollar of spend”, and log transformations are how they get them. In a log-log model, where both sales and the input are logged, the coefficient is the elasticity:
So a price coefficient of means a 1 percent price rise goes with roughly a 2 percent fall in volume, which is directly usable in a pricing decision (the price elasticity guide builds a profit-maximizing price from it). Logging media spend also builds in diminishing returns, since each doubling of spend adds the same percentage lift. A log-linear model, with only the outcome logged, gives coefficients that read as percentage changes in sales per unit of the input, which suits promotions and dummy variables. Choosing between these isn’t a matter of taste; it changes the shape of every response curve that comes out of the model.
Identification: can the data tell the effects apart?
The last habit is asking whether the data contains enough independent variation to estimate each effect. If search and social spend always rose and fell together, no amount of statistics can separate them (multicollinearity in MMM covers what to do), and if spend was always increased before busy periods, the model can mistake demand for media effect (confounding variables). An experienced econometrician will often tell you, before fitting anything, which channels the model will struggle with, just by looking at the spend charts.
Why trends fool regressions
The best way to understand the stationarity problem is to watch it happen. The two series below are independent random walks: Each week, each one moves up or down by a random amount, with no connection whatsoever to the other. Regress one on the other in levels, and the result will very often look significant, with an R² that would make you want to put it in a deck. Re-roll the data a few times, then switch to weekly changes and watch the relationship disappear.
This is the result Clive Granger and Paul Newbold published in 1974 in “Spurious regressions in econometrics”. They regressed independent random walks of 50 observations on each other, 100 times, and the standard 5 percent t-test rejected the null hypothesis of no relationship 76 percent of the time; regressing the changes instead brought that down to 8 percent (Granger and Newbold, 1974). The simulation above, with 1,000 pairs, gives a similar picture: roughly two-thirds or more in levels, depending on the length you choose, and close to the correct 5 percent in changes. Their conclusion was blunt: When a regression in levels has a high R² and a low Durbin-Watson statistic, the fit tells you nothing about whether the relationship is real. The warning itself is older still; the statistician Udny Yule asked “Why do we sometimes get nonsense-correlations between time-series?” in a 1926 address to the Royal Statistical Society.
Now add drift. Move the drift slider up and both series pick up a steady upward trend, like a growing company’s revenue and its growing marketing budget. The correlation in levels heads towards 1 and almost every simulated pair looks significant, even though neither series has any effect on the other. That’s the situation in most marketing datasets: Revenue grows, budgets grow with it (often because revenue grew), and a naive regression hands the credit for the company’s growth to whichever channel’s budget grew fastest.
How econometricians handle trending marketing data
Differencing everything isn’t the whole answer, and Granger and Newbold were careful to say so; their suggestion was to estimate models in both levels and changes and interpret the two together. Advertising effects are often slow and cumulative, and adstock deliberately smooths spend into a slowly moving series, so a real effect can look weaker in weekly changes than it is. In practice, marketing econometricians use a mix of approaches. They include an explicit trend term or a baseline that’s allowed to evolve over time (baseline sales explains why that choice moves so much credit), they add the drivers that cause the trend (distribution, price, population, the size of the customer base) rather than leaving them in the error, and they check that the residuals look like noise rather than a slow-moving pattern.
The test I’d apply to any media coefficient is whether it’s identified by changes in spend rather than by the shared trend. A channel whose estimated effect survives a model with a trend term, or shows up when spend jumps and drops within a year, has a case. A channel whose effect only exists in levels, and vanishes the moment you account for the company’s growth, is probably being credited with the trend. The script below runs exactly that screen on any weekly CSV, which is a useful five-minute check before you trust a vendor’s model or your own.
How econometrics relates to modern Bayesian MMM
If you open Meta’s Robyn, Google’s Meridian or PyMC-Marketing, you’ll find the econometric skeleton intact: sales regressed on adstocked, saturated media, plus controls for price, promotions, trend and seasonality. What’s changed is mostly how the coefficients are estimated and how uncertainty is reported. Robyn uses ridge regression with automated hyperparameter search, while Meridian and PyMC-Marketing are Bayesian: They combine the data with prior beliefs about each channel’s effect and return a full distribution for every ROI rather than a single number with a standard error.
The Bayesian approach solves some classic econometric problems well. Priors stabilize channels the data can’t identify on its own, hierarchical models pool information across regions, and experiment results can go straight in as priors (calibrating an MMM with experiments). However, it doesn’t make the time series problems go away. A Bayesian model fitted to two trending series will happily return a confident posterior for a spurious effect, and a prior can hide the fact that the data contains almost no information about a channel. In my experience the best modern MMM work combines the two traditions: Bayesian estimation and calibration from the new tools, and the econometrician’s suspicion of trends, autocorrelated residuals and coefficients that only exist in levels.
That’s also how the Vexpower course teaches it, starting with plain regression on one invented brand’s weekly data and building up to priors and calibration, with the residual checks at every step. Whether your organization calls it econometrics, marketing mix modeling or MMM, the question underneath is the same one Yule asked a century ago: Is this relationship real, or are two things just moving in the same direction at the same time?
