Response Model Lab

Model Fitting

Name the shape, pick the form, then turn the parameters by hand until the error stops falling. This is the exercise regression automates two weeks from now.

👨‍🏫 Professor Mode: Guided Learning Experience

New to response models? Enable Professor Mode for step-by-step guidance through naming a shape, choosing a functional form, and calibrating it by hand!

Overview & Learning Objectives

A response model says how a marketing input turns into a market outcome. Building one is four decisions, and only the last is arithmetic: what shape do you believe the relationship has, which equation can express that shape, what values should its parameters take, and how will you judge one set of values against another.

By the end of this lab you should be able to:

Why do this by hand? Regression does exactly this search, only faster and without the trial and error. Doing it manually once is what stops least squares looking like magic, and it makes clear that the method automates the search while leaving every judgement about shape and form to you.

Marketing Scenarios

Or upload your own data

Two numeric columns: the marketing input first, the outcome second. A header row is expected and is used to label the axes.

Drop a CSV here or

Step 1Name the shape you expect

Read the scenario above and commit to a shape before you look at the scatter. Choosing the form because it happened to fit best is not a theoretical justification, and it is how overfitting starts.

Steps 2 & 3Pick a form, then calibrate it

Parameters

The automatic search minimises whichever loss is selected above. Run it once under each loss and compare the parameters it lands on.

MAE
average miss
MSE
squared, punishes big misses
Largest miss
single worst point

Where the model is wrong

A loss value is one number for the whole curve. The residual plot shows where the misses are. Structure here — a run of points all above the line, or a U shape — means the form itself is wrong, and no amount of parameter tuning will fix it.

Compare the three forms

Each form is fitted to the current scenario under the selected loss. Fit is not the only criterion: a form that fits marginally better but cannot respect a ceiling, or that bends violently outside the observed range, is usually the worse choice.

Exploratory Questions

  1. Fit the TV reach scenario with a linear model, then read the prediction at 2,000 GRPs. What does it claim, and why is that impossible?
  2. On the email frequency scenario, fit a linear model and then a power series. What does the linear model recommend about send frequency, and what does it miss?
  3. Pick any scenario. Fit it once minimising MAE and once minimising MSE. Where do the two curves differ most, and which points explain the difference?
  4. Add an extreme point to a scenario by uploading a modified CSV. Which loss function moves further, and what does that tell you about using MSE when your data contains genuine outliers?
  5. The trade promotion scenario keeps rising across the observed range. Would you still choose a form with a ceiling? What would you need to know to decide?