Course · 5. TURF (reach)
TURF: what to expect
TURF in Chosŏn is a standard method, offered as a utility next to the menu simulator. There is no claim that it finds better line-ups than other software.
What is standard
- The search. For the sizes people use, an exact solver finds the best set in about a second. Stepwise searches, adding the item that reaches the most new people, come within a fraction of a point. In our tests on real basket data with 30-100 items, stepwise with swaps was within 0.04 points of the best set.
- The sets it picks. Against treating every visit as a new person, the menu model gives the same best sets in the tutorial, or sets tied within noise. On real basket data, standard hierarchical Bayes TURF (each person with their own tastes, items unrelated) picked the same sets as a correlated model like ours at every size, and its held-out reach error was 0.13 points higher (95% range -0.06 to +0.32): a tie.
Where the menu model changes the answer
- Reach over several visits. A model that treats each visit as a new person overstates how many different customers a set reaches over a month: by about 10 points for one item over 4 visits in the tutorial. That number matters for sizing a promotion or a loyalty offer. Models that give each person their own lasting tastes, as ours does, do not have this problem.
Not yet
- TURF from ratings, and from MaxDiff with a real "would use" threshold (anchored MaxDiff).
- Prices other than the base prices in the reach search (use the simulator for a set at other prices).
- Costs and margins.
End of chapter 5. Back to the course contents.