Free for classrooms
Teach cause and effect with the answers known.
Students can't check a causal analysis on real data, because nobody knows the true answer. In these 20,000 simulated worlds we do. Each world was simulated twice, with and without a change. Students score before-and-after, difference-in-differences and a pretrained causal model on error, coverage and false alarms, and see for themselves why most "it worked" charts are wrong.
What's in the lab
Graded exercises
Look at a world and its hidden twin, write a scoring function, then score three methods on 300 worlds. Discussion and project questions at the end.
Before vs after fails
On the lab's worlds, a textbook before-and-after 90% range holds only about 1 time in 5, and claims an effect in about 9 of 10 worlds where nothing happened.
Open data
counterfact-worlds: 20,000 worlds with confounders, seasons, festivals, shocks and spillovers. 25% have no effect at all.
A baseline to beat
counterfact-v2 runs on a laptop. Class project: build a method that beats its false-alarm rate without losing coverage.
Good for
Econometrics and causal inference courses, MBA marketing-analytics modules, data-science bootcamps, public-policy evaluation training, and journalism data courses.
- Students open the notebook in Colab. Nothing to install, nothing to sign up for.
- Instructors can ask for worked solutions and free classroom access to the web app and the Google Sheets formula.
- If you use it in a course or a paper, cite the dataset (BibTeX on the dataset page).