Skip to content

Teaching notes · lesson 11 · a 20-minute class

Marketing mix modelling

What a model can see in your sales data, what it can't, and how sure to be. Tracks: Media.

Open the lesson

Learning goalsBy the end, students can

  • Explain what a marketing mix model is and what it learns from.
  • Read a credible range and say whether a channel is known to pay back.
  • Name three ways to sharpen a model: varied spend, priors and lift tests.

The 20-minute classMinute by minute, with the lab on the projector

  1. 0:003 min, to 3:00

    Cold open

    Ask: last year we spent the same on TV every week. Can a model tell us what TV did? Take a vote.

  2. 3:004 min, to 7:00

    The idea

    Describe the regression in words: weekly sales against media, price, deals and season. Introduce carry-over, saturation and a credible range.

  3. 7:006 min, to 13:00

    Lab on the projector

    Fit with flat spend, then varied spend. Then add timing to the season, then a prior, then a lift test. Watch the range against the true return.

  4. 13:002 min, to 15:00

    Pairs

    Pairs say which change narrowed the range most, and why.

  5. 15:003 min, to 18:00

    Discussion

    Take the range that crosses break-even. What would you do next?

  6. 18:002 min, to 20:00

    Check

    Three lesson questions.

Discussion questionsWith what a strong answer covers

  1. A channel's 80% range for gross profit per €1 is €0.70 to €1.60. Do you keep spending?

    A strong answer covers
    • The range crosses €1, so a loss is still plausible.
    • Run a lift test or vary spend on purpose to narrow it.
    • Avoids treating the point estimate as fact.

  2. Why does spend timed to the season mislead a plain model?

    A strong answer covers
    • Sales are high anyway in season, and spend is high at the same time.
    • The model hands the season's sales to media.
    • A season control or an experiment separates them.

  3. What can't a mix model see?

    A strong answer covers
    • Creative quality: it sees euros, not ads.
    • Most of the long run when fitted on weekly data.
    • It can't tell apart channels that always ran together.

The lab as homeworkAbout 15 minutes for a student

Use the lab to find the cheapest way to make a model's estimate trustworthy.

  1. Fit with flat spend and no prior; note the range.
  2. Change one setting at a time: varied spend, a prior, a lift test.
  3. Note how much each narrows the range and whether it covers the true return.

They answer: Which single change narrowed the range most, did the range still cover the true return, and what would you recommend a brand do first?

A strong answer covers
  • Quotes the ranges from the lab.
  • Checks the range against the true return the lab shows.
  • Recommends deliberate variation or a lift test, with a reason.

Students also answer the lesson's three questions. Their score is worked out on the server from their answers.

Set this lab as homework

Students open a link, do the lab, answer the three questions and write what they found. You see each student's quiz score and finding on your class page.

No classes in this browser yet. Create a class first, then come back.