Upload your choice survey. Chosŏn fits a model that learns, from the choices themselves, which products people treat as substitutes, then shows the results with 95% ranges and lets you simulate the market.
Four kinds of study on one platform. Each comes with a tutorial project and a guided tour, so you can see the full workflow on example data before you upload your own.
Which product people pick from a set, what each feature and price is worth, and where share goes when the line-up changes.
Build-your-own menus and add-ons: what people buy together, how each price moves every item, and bundles.
Rank a long list of features, claims or messages by preference, with a design generator for balanced sets.
The shortlist that reaches the most people, from a menu study or the top of each person's MaxDiff ranking.
The workflow from upload to decision, without scripting.
CSV in a long format, or a Sawtooth .CHO export. Plain-language messages for anything that won't fit, before you spend a run on it.
Feature and price values, shares and reach, each with a range, and a holdout check of how well the model predicts choices it did not see.
Change a product, a price or the whole line-up and see shares, revenue and profit move. Let the optimizer search for the best line-up.
Groups of respondents with different tastes, how distinct the groups are, and how they differ on your profile questions.
Per-respondent estimates as CSV for your own analysis.
Short chapters on conjoint, menus, MaxDiff and TURF, each linked to the screen it explains.
Standard conjoint software estimates what each respondent likes, then decides how people switch with a fixed rule: products that share features are treated as similar. Chosŏn learns switching from the survey instead.
Chosŏn's ranges hold the true answer for every source. The standard method's miss three of the five: too few buyers from the other berry drink, too many from Apex Citrus and from people who would otherwise buy nothing.
Share of the new drink's sales taken from each product. Dots are each method's estimate; lines are the range it is 95% sure of. The shoppers were simulated, so the true answer is known.
| Sales from | True | Chosŏn | Standard method |
|---|---|---|---|
| Surge Berry | 42% | 38% (31–44) | 34% (30–39) |
| Apex Citrus | 25% | 28% (23–33) | 31% (26–36) |
| Volt Original | 19% | 20% (17–30) | 19% (15–23) |
| No purchase | 10% | 10% (8–13) | 12% (10–14) |
| Nova Mango | 4% | 3% (2–4) | 3% (3–5) |
When a product is missing, the model knows where its buyers go: the same flavor, the same pack, the same brand, or anywhere. You don't draw a decision tree or declare which features count before the study.
Line extensions, flavors and pack sizes that people see as close substitutes share demand, so adding a variant doesn't inflate the line's total.
Go / no-go decisions depend on the level of the forecast, not only which option ranks first. In our tests the level lands within about 3% of the truth.
Test the shelf as people see it, 20 to 60 products at once, with switching learned across all of them, in minutes.
Scored on studies where the right answer is known, so every pick and every forecast can be checked. Have a study of your own? Become a beta tester and see what Chosŏn finds in it.
Where shoppers see some products as closer substitutes than others, learning how they switch changes the answer.
Will a new flavor or size grow the brand, or take its sales from your bestseller?
The standard method gave up 1.7–2.7% of profit; Chosŏn picked the best range every time. A ranging test where the right answer is known.
Pick the one that fits. Every route includes all four study types, the simulator, the optimizer and the course.
Homework-size studies, for as long as you like.
30 days with a few larger runs, to try it on a real study.
The trial's larger runs, with a .edu email address.
A real problem and real data, in return for your feedback. Includes our largest runs.
The survey's design and each respondent's choices: one row per product shown, with the respondent, the task, the features and whether it was chosen. Sawtooth .CHO exports load directly. The tutorial projects show the format, and the upload screen tells you what's missing.
For choice-based conjoint: up to 300 respondents, 12 tasks each, 5 products per task and 25 estimated values (what's free, in full). Menu studies up to 300 respondents, 12 menus and 12 items; MaxDiff up to 300 respondents, 20 sets of up to 6, and 30 items. That covers class projects and small pilots. Larger studies run on the trial, academic and beta routes.
A probit choice model with random tastes and correlated errors. The correlations are what carry switching: two products whose errors move together are substitutes. Estimating this at shelf size used to take hours; Chosŏn does it in minutes. The course explains it without the algebra.
No. Each study type has a tutorial project with a guided tour that walks through the results screens. The course is there if you want the background.
Your projects are visible only to your account. To have a project and its data deleted, email info@quantecarlo.com.
A small optimization company in Claverack, New York. Chosŏn is its conjoint platform.