Marketing Mix Modeling Lab

Upload data, validate, fit a model, read ROI with its uncertainty, optimize the budget, export the plan.

Free
Runs in your browser
No sign-up
Synthetic example included

Your weekly data

One row per week: a date, the KPI you want to explain, spend per paid channel, and optional controls such as promotions or holidays.

Upload a CSVStays in your browser. Nothing is uploaded to a server.

No data yet

Upload a weekly CSV or use the synthetic example to walk through the whole workflow.

Build it yourself in Python

The same model, step by step, in a notebook you can run and put in a portfolio.

The notebook builds the model from scratch with pandas and NumPy only: adstock, Hill saturation, ridge regression with non-negative channel effects, the settings search on a validation window, a test window, a bootstrap for ranges, next-dollar returns, and a check against the answer key. It runs on the same synthetic CSV in about a minute.

Download the notebook (.ipynb)Download the data (.csv)

Glossary

Every term the lab uses, in one place.

Baseline: What the KPI would have been with no paid media: trend, seasonality, promotions and other controls.
Incremental: Extra KPI caused by a channel, on top of the baseline.
ROI: Incremental KPI divided by spend, over the whole period. 1.5 means 1.5 back for every 1 spent, before costs.
Next-dollar return (marginal ROI): What one more unit of spend returns at your recent weekly spend. It drives budget decisions. Above the average return means room to grow: the channel is still on the steep, early part of its curve.
Break-even ROI: 1 divided by your margin. At a 40% margin a channel must return 2.5 to pay for itself.
Adstock (carryover): The share of an ad's effect that lands in later weeks. Decay 0.6 means 60% of this week's pressure carries to next week.
Saturation (Hill curve): Diminishing returns: each extra unit of spend does a little less. Half-saturation is where the channel reaches half its maximum effect.
Error on unseen weeks (holdout MAPE): Average % miss on the final weeks, which played no part in fitting. Lower is better; under 10% is good for weekly data.
R²: Share of week to week movement the model explains in the data it was fitted on.
Durbin-Watson: Checks whether errors run in streaks. Near 2 is good; far from 2 suggests a missing driver.
90% range (bootstrap): Four independent searches, each refitted 40 times on resampled data, widened by 1.5 because raw bootstrap ranges proved too narrow. In synthetic tests it held the true ROI in 18 of 18 cases and the true next-dollar return for 21 of 25 channels within their spend history, and 12 of 12 on data built to break its assumptions.
Average return at recent spend: Response divided by spend at your recent weekly level. Compare it with next-dollar return: the whole-period ROI also reflects older, different spend levels.
Confidence grade: About the model, not about money: does it predict unseen weeks, do independent runs agree, are channel ranges tight. Whether marketing pays back is the profit line.
Collinear: Two channels whose spend rises and falls together. No model can split credit between them reliably.
Extrapolation: A prediction for spend levels beyond anything in your history. The curve there is a guess.
Geo / holdout test: Change spend in some regions or for some users and not others, then compare. The way to confirm cause for one channel.

How it works

What this model does

A marketing mix model explains a weekly KPI, such as revenue, as a baseline (trend, seasonality, promotions, holidays) plus the effect of each paid channel. Two transforms make the channel effects realistic. Adstock carries a share of each week's spend into the following weeks, because an ad seen today can sell next week. A Hill saturation curve makes each extra dollar return a little less than the one before.

How it is fitted

This lab follows the approach of Meta's open-source Robyn. Four independent searches each try 300 combinations of carryover and saturation settings per channel, fit each with ridge regression (channel effects cannot go negative), and score them on a validation window. A final test window of the last weeks is kept apart and only used to report the error on unseen weeks. A small penalty favours models whose split of credit is not wildly out of line with the split of spend, because fit alone cannot tell a saturated channel from extra baseline. That penalty is a nudge, not a rule.

How sure is it?

The ranges come from a bootstrap. Each search refits its best models on resampled data 40 times, searching the carryover and saturation settings again each time, and the four searches are pooled, so a rerun gives nearly the same answer. Raw bootstrap ranges proved too narrow, so they are widened by 1.5. On synthetic datasets with a known true answer, the 90% range then held the true ROI in 18 of 18 cases and the true next-dollar return for 21 of 25 channels whose spend stayed within their history (12 of 12 on data built to break its assumptions). A budget move is only recommended when the models and all four searches agree and even the cautious gain is positive. Across eleven test datasets, including five with a deliberately wrong budget, no Grow or Trim call contradicted the truth and no recommended move lost revenue; the lab held back on 8 of 23 clear cases rather than guess. On data built to break its assumptions (TV whose effect peaks weeks later, channels that move together, a promotion it cannot see) it made 4 calls, all right, and held back on 4 of 8 clear cases.

What it cannot do

It is not a Bayesian model like Google's Meridian or PyMC-Marketing, which need Python and minutes to hours of compute. Lift test results can calibrate it (under Data, Advanced): the search then prefers models that reproduce your test, which in synthetic tests narrowed that channel's next-dollar range on every seed. It does not separate geography, and it treats every channel's spend as one weekly number. Any MMM is correlational at heart: when two channels always move together, no model can split their credit. A geo or holdout test is how you settle that for one channel.

Learn the pieces first

The regression, correlation and confidence interval ideas behind this lab are covered in the free Analytics from Scratch course. To try them on their own, use the correlation calculator and the confidence interval calculator. To build the same model yourself, download the Python notebook and the synthetic data.

Marketing science is an analyst job

Marketing, data and business analyst roles, scraped daily from company career pages.

Browse analyst jobs