Course · 3. Menu-based choice (MBC)

Reading the menu results

Overview

The main findings in words: the best seller, the most price-sensitive item, the pairs that go together and the pairs that compete, the item with the most loyal buyers, how much of the links between items is about the person rather than the occasion, and for each bundle, how its price moves its items bought separately (with 95% ranges).

Items and prices

Buy rate: the share of visits that include the item, at base prices, next to the rate seen in the data.
Buy again next visit: of people who buy it on one visit, the share who buy it on the next.
Person share: how much of the buy / don't-buy decision is the person's lasting taste rather than the day.
Price sensitivity: the % change in an item's buy rate for a 1% higher price, with a 95% range. The cross-price table shows how each item responds to the others' prices; most cross effects are small and many ranges include zero.

Bought together

Three views of every pair of items:

Same visit: how much more (or less) often two items are bought together than if they were unrelated. Negative means people take one or the other.
Next visit: people who buy the row item today, and how much more often than average they buy the column item next time. Only lasting tastes carry over, so this is what loyalty schemes, retargeting and cross-promotions can use.
Person vs occasion: for each linked pair, how much of the link is the person.

Every number here has a 95% range: price sensitivities from the usual formula, the rest from 200 sets of model parameters drawn around the fit. Hover a heatmap cell for its range. On 20 simulated surveys the ranges contained the truth 92-94% of the time, a little less often than 95%. The person / occasion share of a single pair is wide with a few hundred respondents (often most of 0 to 100%); the overall share is much tighter (31% to 45% in the tutorial).

Diagnostics

The holdout check compares predictions with each respondent's held-out last menu: errors in percentage points for buy rates, same-visit pairs and pairs from the previous menu to the last one, for both models, with the difference and its 95% range. The observed rates are noisy themselves, so the page also shows the noise floor: the error a perfect model would still have. In the tutorial the two models tie on single visits and ours is closer on the next-visit pairs: 1.07 points against 1.49, a difference of 0.42 [0.32, 0.71].

Try it in the tutorial: who comes back for what

Next: The menu simulator

Tutorial links open the example project in the app; it is free and needs a free account.