PulseGym loses members every month, and the manager wants to know who is at risk before they leave, so the team can reach out in time. Your job: predict which memberships will be cancelled, and explain why, in words a gym manager understands.
The data (in data/)
Real-world messy, on purpose. It covers January 2025 to September 30, 2026 (the cutoff date). Nothing after the cutoff is included.
| File | What's in it |
|---|---|
members.csv |
One row per member: gender, birth date, join date, distance from home, student or not, how they signed up |
memberships.csv |
Every membership: plan, start and end dates, price, payment method |
attendance.csv |
Every check-in: time, duration, activity |
payments.csv |
Every payment attempt: date, amount, paid or failed |
support_tickets.csv |
Complaints and their topic |
train_labels.csv |
Past memberships with the known answer: cancelled = 1 or 0 |
scoring_list.csv |
The memberships you must predict: active on the cutoff date, ending between Oct 1 and Nov 30, 2026 |
Expect missing values, duplicate rows, mixed date formats, inconsistent text (for example Monthly, monthly, MONTHLY), and a few impossible values. Cleaning is part of the job.
The definition
A membership counts as cancelled if the member does not start a new membership within 30 days after its end date.
What you do
- Clean the data, and explain every cleaning decision.
- Explore it and find what drives cancellations, with clear charts.
- Build a prediction model and measure it properly: split your data into training and validation, and never test on the data you trained on.
- No time travel: for each training example, use only data that would have been known before you made the prediction. Your model will predict the scoring list using data up to September 30 only, so train it the same way.
- Name the top reasons members leave, in plain words a gym manager would understand.
What you submit
predictions.csvwith exactly these columns, one row per membership inscoring_list.csv:
membership_id,will_cancel,probability
S207481,1,0.82
S207502,0,0.11
will_cancel is 0 or 1. probability is between 0 and 1.
- Your notebook or code, so we can re-run it.
- A short report (2 pages max): cleaning decisions, what you found, how you validated the model, the top reasons members leave, and how you used AI tools.
How you are scored (out of 100, same sheet for all tracks)
| Criteria | Points |
|---|---|
| Passes the hidden test | 35 |
| Quality of the approach | 25 |
| Edge cases and thinking ahead | 15 |
| Clear explanation (report) | 15 |
| Presentation (easy to run and read) | 10 |
The hidden test: we know what actually happened to every membership in the scoring list. Your predictions are scored with F1 on the cancelled class; we also look at precision, recall and the confusion matrix.
We reward the most reliable, well-validated solution, not the most complex model. A simple model with honest validation beats a fancy one that leaks future data.